Promote clean full-SFT v2 epoch 3 step 5982

README.md

@@ -2,9 +2,11 @@
 license: apache-2.0
 language:
 - en
+- fr
 pipeline_tag: text-generation
-library_name: pytorch
+library_name: transformers
 tags:
+- pytorch
 - safetensors
 - tr-hash
 - mixture-of-experts
@@ -14,64 +16,68 @@ tags:
 - custom-code
 base_model: AETHORIA-AI/TR-HASH-MoE-200M-160B-Refinement
 datasets:
-- AETHORIA-AI/luciole-16way-sft-209k
+- AETHORIA-AI/TR-HASH-MoE-200M-SFT-v2-300K
 ---
 
-# TR-HASH MoE 200M — 160B-source Full SFT
+# TR-HASH MoE 200M — 160B-source Full SFT v2
 
 Full-parameter instruction SFT of
-[`AETHORIA-AI/TR-HASH-MoE-200M-160B-Refinement`](https://huggingface.co/AETHORIA-AI/TR-HASH-MoE-200M-160B-Refinement)
-on the audited 209k-example Luciole 16-way mixture. This is **not LoRA or QLoRA**:
-all 201,194,368 parameters were trainable.
+[`AETHORIA-AI/TR-HASH-MoE-200M-160B-Refinement`](https://huggingface.co/AETHORIA-AI/TR-HASH-MoE-200M-160B-Refinement) on the audited
+[`AETHORIA-AI/TR-HASH-MoE-200M-SFT-v2-300K`](https://huggingface.co/datasets/AETHORIA-AI/TR-HASH-MoE-200M-SFT-v2-300K). This release is
+**not LoRA or QLoRA**: all 201.2M model parameters were trainable.
 
-The root `model.safetensors` is the PIQA-selected **epoch 2 / step 926** model.
-All three resumable training checkpoints remain in `step_000463`, `step_000926`,
-and `step_001389`.
+The root `model.safetensors` is the **epoch 3 / step
+5,982** checkpoint. The strict behavioral regression gate did not pass; this checkpoint was manually promoted because it led all three epochs on full PIQA and held-out SFT loss. The failed checks remain published for transparency.
 
 ## Results
 
-| Epoch | Step | Held-out SFT loss | SFT ppl | PIQA acc | PIQA acc_norm |
-|---:|---:|---:|---:|---:|---:|
-| 1 | 463 | 1.260493 | 3.53 | 68.55% | 69.31% |
-| 2 | 926 | 1.226242 | 3.41 | 68.82% | 69.31% |
-| 3 | 1,389 | 1.220861 | 3.39 | 68.55% | 69.31% |
+| Epoch | Step | Held-out SFT loss | SFT ppl | PIQA acc | PIQA acc_norm | Behavior gate |
+|---:|---:|---:|---:|---:|---:|:---:|
+| 1 | 1,994 | 0.990943 | 2.69 | 67.90% | 68.93% | fail |
+| 2 | 3,988 | 0.963912 | 2.62 | 67.85% | 68.82% | fail |
+| 3 | 5,982 | 0.959617 | 2.61 | 68.01% | 69.10% | fail |
 
-Epoch 2 was promoted because all epochs tie on PIQA `acc_norm`, while epoch 2
-has the highest raw PIQA accuracy. Epoch 3 has the lowest held-out SFT loss but
-was not silently substituted for the benchmark-selected release.
-
-PIQA protocol: full 1,838-example validation split, zero-shot causal
-continuation log-likelihood, no chat template, maximum sequence length 2,048,
-FP16 eager PyTorch with the custom Triton kernels enabled.
+PIQA uses the complete 1,838-example validation split, zero-shot causal
+continuation log-likelihood, no chat template, maximum length 2,048 and FP16
+PyTorch with the custom Triton path. Behavioral reports and exact prompts are
+published under `training/sft-v2-300k/evaluations/`.
 
 ## Training recipe
 
 | Setting | Value |
 |---|---|
 | Method | Full-parameter supervised fine-tuning |
-| Dataset | Luciole 16-way, 209,000 train / 2,100 held out |
-| Supervision | Final assistant response only; prior assistant turns masked |
+| Source | Refinement step 8,156 (about 162B prior token exposures) |
+| Dataset | 300,000 train / 3,000 held-out examples |
+| Tokenized corpus | 202,948,693 train tokens; no truncation |
+| Supervision | Final assistant turn only; prior assistant turns masked |
 | Epochs | 3 |
-| Sequence cap | 512 tokens |
-| Tokenizer | TR-HASH 32k vocabulary |
+| Context | 2,048 tokens |
+| Tokenizer | TR-HASH 32,000-token vocabulary; EOS `</s>` (ID 0) |
 | Optimizer | AdamW, betas 0.9 / 0.95, weight decay 0.1 |
-| LR schedule | 2e-5 peak, 3% warmup, one continuous cosine decay over all epochs |
+| LR | 2e-5 peak, 3% warmup, continuous cosine decay |
 | Precision | BF16 training |
+| Root SafeTensors precision | bfloat16 |
 | Kernels | Liger required; custom Triton enabled |
 
-The tokenized dataset view is published under
-`tokenized/tr-hash-32k-v1/` in
-[`AETHORIA-AI/luciole-16way-sft-209k`](https://huggingface.co/datasets/AETHORIA-AI/luciole-16way-sft-209k),
-including hashes and the assistant-only label mask.
+## Architecture and loading
+
+201.2M parameters, 16 decoder layers, GQA (14 query heads / 2 KV heads),
+four stored deterministic token-ID-routed experts with top-2 activation, an
+always-on shared SwiGLU path and tied embeddings. The persisted multi-hash
+routing tables are part of the checkpoint.
+
+The repository includes an autonomous Transformers adapter. Load it with:
 
-## Architecture
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
-201.2M parameters, 16 transformer layers, GQA (14 query heads / 2 KV heads),
-deterministic token-ID-routed TR-Hash MoE (4 stored experts, top-2 active), and
-tied embeddings. Loading requires the Complexity Framework / TR-Hash runtime.
+repo = "AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT"
+tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)
+```
 
 ## License
 
-Apache License 2.0. This license applies to the model files and
-repository-authored material. Source datasets and individual documents retain
-their own licenses and terms. See [`LICENSE`](LICENSE).
+The released SFT checkpoint is Apache-2.0. Source datasets retain their own
+licenses and terms; see the dataset manifest for the per-source audit.

chat_template.json

@@ -9,7 +9,7 @@
   },
   "assistant_only_loss": true,
   "assistant_prefix": "Assistant:\n",
-  "eos_token": "<|endoftext|>",
+  "eos_token": "</s>",
   "id": "complexity-chat-v2",
   "system_format": "System:\n{content}\n\n",
   "system_prompt": "",

config.json

@@ -6,6 +6,11 @@
   ],
   "attention_dropout": 0.0,
   "attention_type": "gqa",
+  "auto_map": {
+    "AutoConfig": "configuration_tr_hash_moe.TRHashConfig",
+    "AutoModelForCausalLM": "modeling_tr_hash_moe.TRHashForCausalLM"
+  },
+  "bos_token_id": 2,
   "causal_context_fusion_size": 0,
   "causal_context_gate_init": 1.0,
   "causal_contextual_mix_init": 0.0,
@@ -20,6 +25,7 @@
   "causal_stable_delta": false,
   "causal_state_rank": 16,
   "collect_moe_telemetry": false,
+  "eos_token_id": 0,
   "expert_initialization": "gpt_normal",
   "extra_config": {},
   "hash_channel_scale_init": 0.0,
@@ -29,6 +35,8 @@
   "initializer_range": 0.02,
   "intermediate_size": 256,
   "is_causal": true,
+  "is_decoder": true,
+  "is_encoder_decoder": false,
   "learn_hash_channel_modulation": false,
   "learn_hash_pair_gates": false,
   "lexical_gqa_gate_init": 0.0,
@@ -51,15 +59,16 @@
   "micro_expert_width": 16,
   "micro_num_experts": 4,
   "mlp_type": "tr_hash_engine",
-  "model_type": "tr_hash",
+  "model_type": "tr_hash_moe",
   "mup_base_width": 256,
   "norm_eps": 1e-06,
   "norm_type": "rmsnorm",
   "num_attention_heads": 14,
   "num_experts": 4,
+  "num_experts_per_tok": 2,
   "num_hidden_layers": 16,
   "num_key_value_heads": 2,
-  "parameter_count": 201194368,
+  "pad_token_id": 1,
   "rope_fraction": 1.0,
   "rope_theta": 10000.0,
   "rope_type": "standard",
@@ -83,7 +92,7 @@
   "token_frequencies": null,
   "top_k": 2,
   "top_k_primary_weight": 0.5,
-  "torch_dtype": "float32",
+  "torch_dtype": "bfloat16",
   "tr_mha_adapter_gate_init": 0.1,
   "tr_mha_adapter_rank": 8,
   "tr_mha_id_other_logit": -2.0,
@@ -94,6 +103,7 @@
   "tr_mha_top_k": 2,
   "tr_mha_verifier_gate_init": 0.1,
   "tr_mha_verifier_temperature": 1.0,
+  "unk_token_id": 3,
   "use_cache": true,
   "use_cggr": "auto",
   "use_custom_kernels": true,

configuration_tr_hash_moe.py

@@ -0,0 +1,114 @@
+"""Hugging Face configuration for TR-HASH deterministic MoE models.
+
+This file is intentionally self-contained so it can be copied to the root of
+a Hub model repository and loaded with ``trust_remote_code=True``.
+"""
+
+from __future__ import annotations
+
+from transformers import PretrainedConfig
+
+
+class TRHashConfig(PretrainedConfig):
+    """Configuration for the public TR-HASH decoder-only checkpoints."""
+
+    model_type = "tr_hash_moe"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        hidden_size: int = 896,
+        num_hidden_layers: int = 16,
+        num_attention_heads: int = 14,
+        num_key_value_heads: int = 2,
+        intermediate_size: int = 256,
+        shared_intermediate_size: int = 3072,
+        vocab_size: int = 32000,
+        max_position_embeddings: int = 2048,
+        attention_type: str = "gqa",
+        mlp_type: str = "tr_hash_engine",
+        num_experts: int = 4,
+        num_experts_per_tok: int = 2,
+        top_k_primary_weight: float | None = 0.5,
+        routing_strategy: str = "token_id_multi_hash",
+        route_hash_count: int = 2,
+        shared_expert: bool = True,
+        shared_output_scale: float = 1.0,
+        routed_output_scale: float = 2.0,
+        use_qk_norm: bool = True,
+        norm_eps: float = 1e-6,
+        rope_theta: float = 10000.0,
+        attention_dropout: float = 0.0,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = True,
+        initializer_range: float = 0.02,
+        **kwargs,
+    ) -> None:
+        # Accept native pre-adapter configs while keeping architectural top-k
+        # separate from Transformers' generation ``top_k`` sampling option.
+        legacy_top_k = kwargs.pop("top_k", None)
+        if legacy_top_k is not None:
+            num_experts_per_tok = int(legacy_top_k)
+        if attention_type != "gqa":
+            raise ValueError("The public TR-HASH adapter currently supports GQA only")
+        if mlp_type not in {"tr_hash_engine", "tr_hash_moe"}:
+            raise ValueError("TRHashConfig requires mlp_type='tr_hash_engine'")
+        if hidden_size % num_attention_heads:
+            raise ValueError("hidden_size must be divisible by num_attention_heads")
+        if num_attention_heads % num_key_value_heads:
+            raise ValueError("num_attention_heads must be divisible by num_key_value_heads")
+        if intermediate_size % num_experts:
+            raise ValueError("intermediate_size must be divisible by num_experts")
+        if not 1 <= num_experts_per_tok <= num_experts:
+            raise ValueError("num_experts_per_tok must be between 1 and num_experts")
+
+        self.hidden_size = int(hidden_size)
+        self.num_hidden_layers = int(num_hidden_layers)
+        self.num_attention_heads = int(num_attention_heads)
+        self.num_key_value_heads = int(num_key_value_heads)
+        self.intermediate_size = int(intermediate_size)
+        self.shared_intermediate_size = int(shared_intermediate_size)
+        self.vocab_size = int(vocab_size)
+        self.max_position_embeddings = int(max_position_embeddings)
+        self.attention_type = attention_type
+        self.mlp_type = mlp_type
+        self.num_experts = int(num_experts)
+        self.num_experts_per_tok = int(num_experts_per_tok)
+        self.top_k_primary_weight = top_k_primary_weight
+        self.routing_strategy = routing_strategy
+        self.route_hash_count = int(route_hash_count)
+        self.shared_expert = bool(shared_expert)
+        self.shared_output_scale = float(shared_output_scale)
+        self.routed_output_scale = float(routed_output_scale)
+        self.use_qk_norm = bool(use_qk_norm)
+        self.norm_eps = float(norm_eps)
+        self.rope_theta = float(rope_theta)
+        self.attention_dropout = float(attention_dropout)
+        self.initializer_range = float(initializer_range)
+        self.use_cache = bool(use_cache)
+        self.head_dim = self.hidden_size // self.num_attention_heads
+        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads
+        self.expert_width = self.intermediate_size // self.num_experts
+
+        kwargs.setdefault("architectures", ["TRHashForCausalLM"])
+        # These are emitted by save_pretrained/export and therefore come back
+        # through **kwargs on reload. Consume them before passing the canonical
+        # values below so repeated save/load cycles stay valid.
+        kwargs.pop("is_decoder", None)
+        kwargs.pop("is_encoder_decoder", None)
+        super().__init__(
+            vocab_size=self.vocab_size,
+            max_position_embeddings=self.max_position_embeddings,
+            use_cache=use_cache,
+            tie_word_embeddings=tie_word_embeddings,
+            is_decoder=True,
+            is_encoder_decoder=False,
+            **kwargs,
+        )
+        # Transformers 5 no longer materializes every model-specific keyword
+        # passed to PretrainedConfig, while Transformers 4 did. Keep the
+        # decoder cache contract explicit across both release lines.
+        self.use_cache = bool(use_cache)
+
+
+__all__ = ["TRHashConfig"]

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2dd5a35624dda4cb267d872f974fd448c9e085004c163b31772515bfd9573caa
-size 813515200
+oid sha256:b5a5e0a4eb94ec9e0bb79ab4fefc8b25b3ce6e33e657276c95f43c79b5cbbed4
+size 411125600

model_config.yaml

@@ -0,0 +1,99 @@
+active_expert_width: null
+active_num_experts: null
+attention_dropout: 0.0
+attention_type: gqa
+causal_context_fusion_size: 0
+causal_context_gate_init: 1.0
+causal_contextual_mix_init: 0.0
+causal_conv_dilation_cycle: 8
+causal_conv_kernel_size: 4
+causal_delta_chunk_size: 512
+causal_delta_collision_normalized: false
+causal_delta_lexical_forge: false
+causal_delta_lexical_values: false
+causal_delta_occurrence_address: false
+causal_delta_timescales: 1
+causal_stable_delta: false
+causal_state_rank: 16
+collect_moe_telemetry: false
+expert_initialization: gpt_normal
+extra_config: {}
+hash_channel_scale_init: 0.0
+hash_pair_gate_init: 0.5
+hidden_act: silu
+hidden_size: 896
+initializer_range: 0.02
+intermediate_size: 256
+is_causal: true
+learn_hash_channel_modulation: false
+learn_hash_pair_gates: false
+lexical_gqa_gate_init: 0.0
+lexical_gqa_rank: 16
+lexical_gqa_use_token_code: true
+lexical_key_gate_init: 0.05
+lexical_object_gate_init: 0.1
+lexical_object_rank: 16
+lexical_zipf_alpha: 0.25
+lexical_zipf_floor: 0.1
+lexical_zipf_mode: uniform
+lexical_zipf_path: null
+lexical_zipf_permutation_seed: 1729
+lsh_bits: 0
+lsh_from_layer: 0
+lsh_routing: false
+lsh_threshold_mode: zero
+max_position_embeddings: 2048
+micro_expert_gate_init: 0.1
+micro_expert_width: 16
+micro_num_experts: 4
+mlp_type: tr_hash_engine
+mup_base_width: 256
+norm_eps: 1.0e-06
+norm_type: rmsnorm
+num_attention_heads: 14
+num_experts: 4
+num_hidden_layers: 16
+num_key_value_heads: 2
+rope_fraction: 1.0
+rope_theta: 10000.0
+rope_type: standard
+route_hash_count: 2
+routed_gate_init: 1.0
+routed_output_scale: 2.0
+routed_output_scale_first_layer: null
+routed_output_scale_last_layer: null
+routing_strategy: token_id_multi_hash
+shared_expert: true
+shared_expert_chunk_tokens: 0
+shared_gate_init: 1.0
+shared_intermediate_size: 3072
+shared_output_scale: 1.0
+shared_output_scale_first_layer: null
+shared_output_scale_last_layer: null
+sliding_window: null
+static_expert_capacity: false
+tie_lexical_object_embeddings: false
+tie_word_embeddings: true
+token_frequencies: null
+top_k: 2
+top_k_primary_weight: 0.5
+tr_mha_adapter_gate_init: 0.1
+tr_mha_adapter_rank: 8
+tr_mha_id_other_logit: -2.0
+tr_mha_id_primary_logit: 2.0
+tr_mha_id_secondary_logit: 1.0
+tr_mha_num_experts: 4
+tr_mha_targets: qv
+tr_mha_top_k: 2
+tr_mha_verifier_gate_init: 0.1
+tr_mha_verifier_temperature: 1.0
+use_cache: true
+use_cggr: auto
+use_custom_kernels: true
+use_mup_attn_scale: false
+use_mup_init: false
+use_mup_output_mult: false
+use_qk_norm: true
+use_sdpa: true
+use_shared_routed_gates: false
+vocab_size: 32000

modeling_tr_hash_moe.py

@@ -0,0 +1,548 @@
+"""Self-contained Transformers implementation of TR-HASH MoE.
+
+The module preserves the native checkpoint tensor names. Its universal
+PyTorch expert path is the numerical reference; optimized serving remains the
+job of TR-Hash-i64 and, once supported, vLLM.
+"""
+
+from __future__ import annotations
+
+import math
+from typing import Any
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import GenerationMixin, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from .configuration_tr_hash_moe import TRHashConfig
+
+try:
+    from transformers.cache_utils import Cache, DynamicCache
+except ImportError:  # Transformers 4.30-4.35: legacy tuple caches only.
+
+    class Cache:  # type: ignore[no-redef]
+        pass
+
+    DynamicCache = None  # type: ignore[assignment,misc]
+
+
+_GenerationBase = object if issubclass(PreTrainedModel, GenerationMixin) else GenerationMixin
+
+
+class TRHashRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        variance = hidden_states.pow(2).mean(dim=-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+        return self.weight * hidden_states
+
+
+def _rotate_half(values: torch.Tensor) -> torch.Tensor:
+    first, second = values.chunk(2, dim=-1)
+    return torch.cat((-second, first), dim=-1)
+
+
+class TRHashRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, max_position_embeddings: int, theta: float) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq)
+        self.max_position_embeddings = max_position_embeddings
+
+    def forward(
+        self, position_ids: torch.LongTensor, dtype: torch.dtype
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        frequencies = torch.einsum("bi,j->bij", position_ids.float(), self.inv_freq.float())
+        embeddings = torch.cat((frequencies, frequencies), dim=-1)
+        return embeddings.cos().to(dtype=dtype), embeddings.sin().to(dtype=dtype)
+
+
+class TRHashAttention(nn.Module):
+    def __init__(self, config: TRHashConfig, layer_idx: int) -> None:
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = config.num_key_value_groups
+        self.head_dim = config.head_dim
+        self.dropout = config.attention_dropout
+
+        kv_size = self.num_key_value_heads * self.head_dim
+        self.k_proj = nn.Linear(config.hidden_size, kv_size, bias=False)
+        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, kv_size, bias=False)
+        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        if config.use_qk_norm:
+            self.q_norm = TRHashRMSNorm(self.head_dim, 1e-6)
+            self.k_norm = TRHashRMSNorm(self.head_dim, 1e-6)
+        else:
+            self.q_norm = None
+            self.k_norm = None
+        self.rotary_emb = TRHashRotaryEmbedding(
+            self.head_dim,
+            config.max_position_embeddings,
+            config.rope_theta,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None,
+        position_ids: torch.LongTensor,
+        past_key_values: Cache | tuple[torch.Tensor, torch.Tensor] | None,
+        use_cache: bool,
+        cache_position: torch.LongTensor | None,
+    ) -> tuple[torch.Tensor, Any]:
+        batch_size, sequence_length, _ = hidden_states.shape
+        # Match complexity-framework's fused K/Q/V projection order exactly.
+        projection = F.linear(
+            hidden_states,
+            torch.cat(
+                (self.k_proj.weight, self.q_proj.weight, self.v_proj.weight),
+                dim=0,
+            ),
+        )
+        kv_size = self.num_key_value_heads * self.head_dim
+        key_states, query_states, value_states = projection.split(
+            (kv_size, self.num_heads * self.head_dim, kv_size), dim=-1
+        )
+        query_states = query_states.view(
+            batch_size, sequence_length, self.num_heads, self.head_dim
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            batch_size, sequence_length, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            batch_size, sequence_length, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+
+        if self.q_norm is not None:
+            query_states = self.q_norm(query_states)
+            key_states = self.k_norm(key_states)
+        cos, sin = self.rotary_emb(position_ids, query_states.dtype)
+        cos = cos.unsqueeze(1)
+        sin = sin.unsqueeze(1)
+        query_states = query_states * cos + _rotate_half(query_states) * sin
+        key_states = key_states * cos + _rotate_half(key_states) * sin
+
+        new_past: Any = None
+        if isinstance(past_key_values, Cache):
+            key_states, value_states = past_key_values.update(
+                key_states,
+                value_states,
+                self.layer_idx,
+                {"cache_position": cache_position},
+            )
+            new_past = past_key_values if use_cache else None
+        elif past_key_values is not None:
+            key_states = torch.cat((past_key_values[0], key_states), dim=2)
+            value_states = torch.cat((past_key_values[1], value_states), dim=2)
+            new_past = (key_states, value_states) if use_cache else None
+        elif use_cache:
+            new_past = (key_states, value_states)
+
+        key_states = key_states.repeat_interleave(self.num_key_value_groups, dim=1)
+        value_states = value_states.repeat_interleave(self.num_key_value_groups, dim=1)
+        query_length = query_states.shape[-2]
+        key_length = key_states.shape[-2]
+        is_causal = attention_mask is None and query_length == key_length
+        attention_output = F.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.dropout if self.training else 0.0,
+            is_causal=is_causal,
+            scale=1.0 / math.sqrt(self.head_dim),
+        )
+        attention_output = attention_output.transpose(1, 2).reshape(batch_size, sequence_length, -1)
+        return self.o_proj(attention_output), new_past
+
+
+class TRHashExpertEngine(nn.Module):
+    """Shared SwiGLU plus fixed top-k experts with native tensor names."""
+
+    def __init__(self, config: TRHashConfig) -> None:
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.vocab_size = config.vocab_size
+        self.shared_output_scale = config.shared_output_scale
+        self.routed_output_scale = config.routed_output_scale
+        if self.top_k == 1:
+            route_weights = (1.0,)
+        else:
+            primary = (
+                1.0 / self.top_k
+                if config.top_k_primary_weight is None
+                else float(config.top_k_primary_weight)
+            )
+            route_weights = (
+                primary,
+                *((1.0 - primary) / (self.top_k - 1) for _ in range(self.top_k - 1)),
+            )
+        # Keep these as Python scalars. Transformers may construct the model
+        # on the meta device while streaming a safetensors checkpoint; a
+        # derived non-persistent tensor buffer can otherwise be materialized
+        # as zeros because there is intentionally no checkpoint key for it.
+        self.route_weights = tuple(float(value) for value in route_weights)
+        # The checkpoint supplies these persisted routing artifacts. Zero
+        # initialization prevents a second, subtly different hash builder from
+        # ever being treated as authoritative.
+        self.register_buffer(
+            "route_table",
+            torch.zeros(self.top_k, self.vocab_size, dtype=torch.long),
+        )
+        self.register_buffer(
+            "fused_route_codes",
+            torch.zeros(self.vocab_size, dtype=torch.uint8),
+        )
+        pair_count = self.num_experts * (self.num_experts - 1) // 2
+        self.register_buffer(
+            "fused_expert_pairs",
+            torch.zeros(pair_count, 2, dtype=torch.int32),
+        )
+        self.expert_gate = nn.Parameter(
+            torch.empty(config.num_experts, config.hidden_size, config.expert_width)
+        )
+        self.expert_up = nn.Parameter(
+            torch.empty(config.num_experts, config.hidden_size, config.expert_width)
+        )
+        self.expert_down = nn.Parameter(
+            torch.empty(config.num_experts, config.expert_width, config.hidden_size)
+        )
+        if config.shared_expert:
+            self.shared_gate = nn.Linear(
+                config.hidden_size, config.shared_intermediate_size, bias=False
+            )
+            self.shared_up = nn.Linear(
+                config.hidden_size, config.shared_intermediate_size, bias=False
+            )
+            self.shared_down = nn.Linear(
+                config.shared_intermediate_size, config.hidden_size, bias=False
+            )
+        else:
+            self.shared_gate = self.shared_up = self.shared_down = None
+
+    @property
+    def experts(self) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Expose the sparse expert block to serving/quantization adapters."""
+
+        return self.expert_gate, self.expert_up, self.expert_down
+
+    def forward(self, hidden_states: torch.Tensor, token_ids: torch.LongTensor) -> torch.Tensor:
+        batch_size, sequence_length, hidden_size = hidden_states.shape
+        flat_states = hidden_states.reshape(-1, hidden_size)
+        if self.shared_gate is None:
+            shared = torch.zeros_like(flat_states)
+        else:
+            shared = self.shared_down(
+                F.silu(self.shared_gate(flat_states)) * self.shared_up(flat_states)
+            )
+        routes = self.route_table[:, token_ids.clamp(0, self.vocab_size - 1)].reshape(
+            self.top_k, -1
+        )
+        route_weights = flat_states.new_tensor(self.route_weights).view(-1, 1)
+        routed = torch.zeros_like(flat_states)
+        for expert_index in range(self.num_experts):
+            token_weight = (routes.eq(expert_index).to(flat_states.dtype) * route_weights).sum(
+                dim=0
+            )
+            active_states = flat_states * token_weight.ne(0).to(flat_states.dtype).unsqueeze(-1)
+            intermediate = F.silu(active_states @ self.expert_gate[expert_index]) * (
+                active_states @ self.expert_up[expert_index]
+            )
+            expert_output = intermediate @ self.expert_down[expert_index]
+            routed.add_(expert_output * token_weight.unsqueeze(-1))
+        output = self.shared_output_scale * shared + self.routed_output_scale * routed
+        return output.view(batch_size, sequence_length, hidden_size)
+
+
+class TRHashMLP(nn.Module):
+    def __init__(self, config: TRHashConfig) -> None:
+        super().__init__()
+        self.engine = TRHashExpertEngine(config)
+
+    @property
+    def experts(self):
+        return self.engine.experts
+
+    def forward(self, hidden_states: torch.Tensor, token_ids: torch.LongTensor):
+        return self.engine(hidden_states, token_ids)
+
+
+class TRHashDecoderLayer(nn.Module):
+    def __init__(self, config: TRHashConfig, layer_idx: int) -> None:
+        super().__init__()
+        self.input_layernorm = TRHashRMSNorm(config.hidden_size, config.norm_eps)
+        self.self_attn = TRHashAttention(config, layer_idx)
+        self.post_attention_layernorm = TRHashRMSNorm(config.hidden_size, config.norm_eps)
+        self.mlp = TRHashMLP(config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        token_ids: torch.LongTensor,
+        attention_mask: torch.Tensor | None,
+        position_ids: torch.LongTensor,
+        past_key_values: Any,
+        use_cache: bool,
+        cache_position: torch.LongTensor | None,
+    ) -> tuple[torch.Tensor, Any]:
+        residual = hidden_states
+        attention_output, new_past = self.self_attn(
+            self.input_layernorm(hidden_states),
+            attention_mask,
+            position_ids,
+            past_key_values,
+            use_cache,
+            cache_position,
+        )
+        hidden_states = residual + attention_output
+        return (
+            hidden_states + self.mlp(self.post_attention_layernorm(hidden_states), token_ids),
+            new_past,
+        )
+
+
+class TRHashForCausalLM(PreTrainedModel, _GenerationBase):
+    config_class = TRHashConfig
+    base_model_prefix = ""
+    main_input_name = "input_ids"
+    _supports_cache_class = True
+    _supports_sdpa = True
+    _no_split_modules = ["TRHashDecoderLayer"]
+
+    def __init__(self, config: TRHashConfig) -> None:
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = nn.ModuleList(
+            TRHashDecoderLayer(config, index) for index in range(config.num_hidden_layers)
+        )
+        self.norm = TRHashRMSNorm(config.hidden_size, config.norm_eps)
+        # Besides initializing newly-created models, post_init records the
+        # loading/tied-weight metadata required by both Transformers 4 and 5.
+        # from_pretrained replaces every persisted tensor immediately after.
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.embed_tokens
+
+    def set_output_embeddings(self, value):
+        self.embed_tokens = value
+
+    def tie_weights(self, *args, **kwargs):
+        return None
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values=None,
+        attention_mask: torch.Tensor | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        **kwargs,
+    ) -> dict[str, Any]:
+        """Bridge the generation contracts used by Transformers 4 and 5."""
+
+        if isinstance(past_key_values, Cache):
+            has_cached_values = past_key_values.get_seq_length() > 0
+        else:
+            has_cached_values = past_key_values is not None
+        if has_cached_values:
+            input_ids = input_ids[:, -1:]
+        position_ids = kwargs.get("position_ids")
+        if position_ids is None and attention_mask is not None:
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask.eq(0), 0)
+        if position_ids is not None:
+            position_ids = position_ids[:, -input_ids.shape[1] :]
+        model_inputs: dict[str, Any]
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+        model_inputs.update(
+            {
+                "attention_mask": attention_mask,
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache", True),
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx: torch.LongTensor):
+        if isinstance(past_key_values, Cache):
+            past_key_values.reorder_cache(beam_idx)
+            return past_key_values
+        return tuple(
+            tuple(state.index_select(0, beam_idx.to(state.device)) for state in layer)
+            for layer in past_key_values
+        )
+
+    @staticmethod
+    def _causal_padding_mask(
+        attention_mask: torch.Tensor | None,
+        batch_size: int,
+        query_length: int,
+        key_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+    ) -> torch.Tensor | None:
+        needs_offset_causal_mask = key_length != query_length and query_length > 1
+        has_padding = attention_mask is not None and not bool(attention_mask.all())
+        if not needs_offset_causal_mask and not has_padding:
+            return None
+        minimum = torch.finfo(dtype).min
+        query_positions = torch.arange(
+            key_length - query_length, key_length, device=device
+        ).unsqueeze(-1)
+        key_positions = torch.arange(key_length, device=device).unsqueeze(0)
+        causal = key_positions > query_positions
+        mask = causal.view(1, 1, query_length, key_length).expand(
+            batch_size, 1, query_length, key_length
+        )
+        if attention_mask is None:
+            padding = torch.zeros_like(mask)
+        else:
+            padding = attention_mask[:, None, None, :key_length].eq(0)
+        return torch.zeros(mask.shape, dtype=dtype, device=device).masked_fill(
+            mask | padding, minimum
+        )
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | tuple | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs,
+    ) -> CausalLMOutputWithPast | tuple:
+        cache_position = kwargs.pop("cache_position", None)
+        if output_attentions:
+            raise NotImplementedError("TR-HASH does not return attention weights")
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("input_ids or inputs_embeds must be provided")
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("Specify only one of input_ids and inputs_embeds")
+        use_cache = self.config.use_cache if use_cache is None else use_cache
+        output_hidden_states = (
+            self.config.output_hidden_states
+            if output_hidden_states is None
+            else output_hidden_states
+        )
+        return_dict = self.config.return_dict if return_dict is None else return_dict
+
+        hidden_states = self.embed_tokens(input_ids) if inputs_embeds is None else inputs_embeds
+        batch_size, query_length = hidden_states.shape[:2]
+        token_ids = input_ids
+        if token_ids is None:
+            token_ids = torch.zeros(
+                batch_size, query_length, dtype=torch.long, device=hidden_states.device
+            )
+
+        if isinstance(past_key_values, Cache):
+            past_length = past_key_values.get_seq_length()
+        elif past_key_values:
+            past_length = past_key_values[0][0].shape[2]
+        else:
+            past_length = 0
+            if use_cache and past_key_values is None and DynamicCache is not None:
+                past_key_values = DynamicCache(config=self.config)
+        if cache_position is None:
+            cache_position = torch.arange(
+                past_length,
+                past_length + query_length,
+                device=hidden_states.device,
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0).expand(batch_size, -1)
+        key_length = past_length + query_length
+        causal_mask = self._causal_padding_mask(
+            attention_mask,
+            batch_size,
+            query_length,
+            key_length,
+            hidden_states.dtype,
+            hidden_states.device,
+        )
+
+        all_hidden_states = [hidden_states] if output_hidden_states else None
+        legacy_cache = [] if use_cache and not isinstance(past_key_values, Cache) else None
+        for layer_index, layer in enumerate(self.layers):
+            layer_past = (
+                past_key_values
+                if isinstance(past_key_values, Cache)
+                else (past_key_values[layer_index] if past_key_values else None)
+            )
+            hidden_states, new_past = layer(
+                hidden_states,
+                token_ids,
+                causal_mask,
+                position_ids,
+                layer_past,
+                use_cache,
+                cache_position,
+            )
+            if legacy_cache is not None:
+                legacy_cache.append(new_past)
+            if all_hidden_states is not None:
+                all_hidden_states.append(hidden_states)
+        hidden_states = self.norm(hidden_states)
+        if all_hidden_states is not None:
+            all_hidden_states[-1] = hidden_states
+
+        if isinstance(logits_to_keep, int) and logits_to_keep > 0:
+            selected = hidden_states[:, -logits_to_keep:, :]
+        elif isinstance(logits_to_keep, torch.Tensor):
+            selected = hidden_states[:, logits_to_keep, :]
+        else:
+            selected = hidden_states
+        logits = F.linear(selected, self.embed_tokens.weight)
+        loss = None
+        if labels is not None:
+            shift_logits = logits[:, :-1].contiguous().float()
+            shift_labels = labels[:, 1:].contiguous()
+            loss = F.cross_entropy(
+                shift_logits.view(-1, self.config.vocab_size),
+                shift_labels.view(-1),
+                ignore_index=-100,
+            )
+
+        returned_cache = past_key_values if isinstance(past_key_values, Cache) else legacy_cache
+        if not return_dict:
+            values = (
+                logits,
+                returned_cache,
+                tuple(all_hidden_states) if all_hidden_states else None,
+            )
+            return ((loss,) + values) if loss is not None else values
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=returned_cache,
+            hidden_states=tuple(all_hidden_states) if all_hidden_states else None,
+        )
+
+
+__all__ = ["TRHashForCausalLM", "TRHashConfig"]

release_manifest.json

@@ -1,13 +1,25 @@
 {
-  "dataset": "AETHORIA-AI/luciole-16way-sft-209k",
+  "architecture": "tr_hash_moe",
+  "behavior_gate_passed": false,
+  "dataset": "AETHORIA-AI/TR-HASH-MoE-200M-SFT-v2-300K",
+  "dataset_revision": "084a658ec47e4ee872f6d67fdbad3602f599424b",
+  "floating_parameters": 201194880,
+  "matched_eval_loss": 0.959617,
+  "matched_eval_ppl": 2.61,
   "method": "full-parameter-sft",
-  "piqa_acc": 0.6882480957562568,
-  "piqa_acc_norm": 0.6931447225244831,
+  "num_experts": 4,
+  "num_experts_per_tok": 2,
+  "parameters": 202731072,
+  "piqa_acc": 0.6800870511425462,
+  "piqa_acc_norm": 0.690968443960827,
   "release": "AETHORIA-AI/TR-HASH-MoE-200M-160B-SFT",
-  "schema_version": 1,
-  "selected_epoch": 2,
-  "selected_step": 926,
-  "selection": "max PIQA acc_norm, then max PIQA acc",
+  "schema_version": 2,
+  "selected_epoch": 3,
+  "selected_step": 5982,
+  "selection_policy": "manual promotion of epoch 3: highest full-PIQA acc_norm and acc, lowest held-out SFT loss; strict behavior failures retained",
+  "source_checkpoint_step": 8156,
   "source_model": "AETHORIA-AI/TR-HASH-MoE-200M-160B-Refinement",
-  "weights_sha256": "2dd5a35624dda4cb267d872f974fd448c9e085004c163b31772515bfd9573caa"
+  "weights_bytes": 411125600,
+  "weights_dtype": "bfloat16",
+  "weights_sha256": "b5a5e0a4eb94ec9e0bb79ab4fefc8b25b3ce6e33e657276c95f43c79b5cbbed4"
 }

reports/sft-v2-300k/evaluations/epoch_01_step_001994/chat.json

@@ -0,0 +1,130 @@
+{
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_001994",
+  "chat_template_applied": true,
+  "runtime": "pytorch",
+  "generation": {
+    "max_tokens": 512,
+    "temperature": 0.0,
+    "top_p": 1.0,
+    "top_k": 0,
+    "repetition_penalty": 1.1,
+    "repetition_context_size": 128
+  },
+  "results": [
+    {
+      "id": "greeting",
+      "prompt": "Hello!",
+      "response": "Hi there! How can I help you?",
+      "repetition": {
+        "words": 7,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "math_multiplication",
+      "prompt": "Calculate 17 \u00d7 23. Give the result and one short calculation.",
+      "response": "To calculate 17 \u00d7 23, we can use the following steps:\n\n1. Start with the given product: 17 \u00d7 23 = 441.\n2. Multiply the first number (17) by the second number (23): 17 \u00d7 23 = 441.\n3. Add the results from step 1: 441 + 441 = 842.\n4. Subtract the result of step 3 from the original product: 842 - 441 = 41.\n5. Divide the difference by 2 to find the answer: 41 \u00f7 2 = 26.5.\n6. Convert the decimal to a percentage: 26.5% = 26.5/100 \u2248 0.265.\n7. Round the result to two decimal places if necessary: 0.265 rounded to two decimal places is approximately 26.55%.\n\nTherefore, the final answer is $\\boxed{0.265}$.",
+      "repetition": {
+        "words": 114,
+        "distinct_2": 0.902655,
+        "repeated_trigrams": 3,
+        "repeated_trigram_ratio": 0.026786
+      }
+    },
+    {
+      "id": "math_word_problem",
+      "prompt": "A box contains 24 red balls and 18 blue balls. Seven balls are removed. How many remain? Answer with the number and one sentence.",
+      "response": "The total number of balls is \\(24 + 18 = 42\\). Remaining balls are \\(42 - 7 = 35\\) balls. \n\nAnswer: 35",
+      "repetition": {
+        "words": 18,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "code_add",
+      "prompt": "Write only executable Python code defining add(a, b), which returns the sum of a and b.",
+      "response": "Here's an example of how you can define `add` in Python using the `+` operator:\n\n```python\ndef add(a, b):\n    return a + b\n```",
+      "repetition": {
+        "words": 22,
+        "distinct_2": 0.952381,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "code_is_prime",
+      "prompt": "Write only executable Python code defining is_prime(n). It must return True exactly when integer n is prime.",
+      "response": "Here's the implementation of the function `is_prime` that checks if a number is prime using the Sieve of Eratosthenes algorithm:\n\n```python\ndef is_prime(n):\n    if not isinstance(n, int) or n <= 1:\n        return False\n    for i in range(2, int(n**0.5) + 1):\n        if n % i == 0:\n            return False\n    return True\n```",
+      "repetition": {
+        "words": 52,
+        "distinct_2": 0.980392,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "memory_name",
+      "prompt": [
+        {
+          "role": "user",
+          "content": "My name is Boris. Please remember it."
+        },
+        {
+          "role": "assistant",
+          "content": "Understood, your name is Boris."
+        },
+        {
+          "role": "user",
+          "content": "What is my name? Answer in one short sentence."
+        }
+      ],
+      "response": "Boris is a fictional character created by the author Boris Kornis.",
+      "repetition": {
+        "words": 11,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "memory_rules",
+      "prompt": [
+        {
+          "role": "user",
+          "content": "For our game, the secret piece is a blue triangle and collecting a star gives 3 points."
+        },
+        {
+          "role": "assistant",
+          "content": "Got it: blue triangle, and each star is worth 3 points."
+        },
+        {
+          "role": "user",
+          "content": "Remind me of the secret piece and the star score."
+        }
+      ],
+      "response": "The secret piece is a blue triangle with a star on its side. Each star is worth 3 points.",
+      "repetition": {
+        "words": 19,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "instruction_three_bullets",
+      "prompt": "Give exactly three bullet points for saving electricity. Use no more than eight words per bullet.",
+      "response": "* **Energy Efficiency**: Invest in energy-efficient appliances and devices to reduce your utility bills.\n* **Reduced Energy Consumption**: Choose energy-efficient lighting, appliances, and heating systems to lower your energy usage.\n* **Cost Savings**: Monitor your energy costs and make adjustments as needed to optimize your energy efficiency.",
+      "repetition": {
+        "words": 46,
+        "distinct_2": 0.888889,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    }
+  ]
+}

reports/sft-v2-300k/evaluations/epoch_01_step_001994/evaluation.log

@@ -0,0 +1,105 @@
+Activated conda/uv virtual environment at /venv/main
+[Tokenizer] Loaded from /workspace/tr-hash-moe-200m-sft-v2-300k/tokenized/tr-hash-32k-v2-2048/tokenizer
+
+[greeting]
+Hi there! How can I help you?
+
+[math_multiplication]
+To calculate 17 × 23, we can use the following steps:
+
+1. Start with the given product: 17 × 23 = 441.
+2. Multiply the first number (17) by the second number (23): 17 × 23 = 441.
+3. Add the results from step 1: 441 + 441 = 842.
+4. Subtract the result of step 3 from the original product: 842 - 441 = 41.
+5. Divide the difference by 2 to find the answer: 41 ÷ 2 = 26.5.
+6. Convert the decimal to a percentage: 26.5% = 26.5/100 ≈ 0.265.
+7. Round the result to two decimal places if necessary: 0.265 rounded to two decimal places is approximately 26.55%.
+
+Therefore, the final answer is $\boxed{0.265}$.
+
+[math_word_problem]
+The total number of balls is \(24 + 18 = 42\). Remaining balls are \(42 - 7 = 35\) balls. 
+
+Answer: 35
+
+[code_add]
+Here's an example of how you can define `add` in Python using the `+` operator:
+
+```python
+def add(a, b):
+    return a + b
+```
+
+[code_is_prime]
+Here's the implementation of the function `is_prime` that checks if a number is prime using the Sieve of Eratosthenes algorithm:
+
+```python
+def is_prime(n):
+    if not isinstance(n, int) or n <= 1:
+        return False
+    for i in range(2, int(n**0.5) + 1):
+        if n % i == 0:
+            return False
+    return True
+```
+
+[memory_name]
+Boris is a fictional character created by the author Boris Kornis.
+
+[memory_rules]
+The secret piece is a blue triangle with a star on its side. Each star is worth 3 points.
+
+[instruction_three_bullets]
+* **Energy Efficiency**: Invest in energy-efficient appliances and devices to reduce your utility bills.
+* **Reduced Energy Consumption**: Choose energy-efficient lighting, appliances, and heating systems to lower your energy usage.
+* **Cost Savings**: Monitor your energy costs and make adjustments as needed to optimize your energy efficiency.
+
+Report saved: artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_01_step_001994/chat.json
+22:13:50 | INFO | [device] backend=cuda device=cuda name=NVIDIA GeForce RTX 5090 matmul=cuBLAS/cuBLASLt distributed=NCCL sdpa=true flash_attention=true custom_triton=true
+[Tokenizer] Loaded from /workspace/tr-hash-moe-200m-sft-v2-300k/tokenized/tr-hash-32k-v2-2048/tokenizer
+scored 512/3,676 choices (118.1/s)
+scored 1,024/3,676 choices (202.5/s)
+scored 1,536/3,676 choices (273.0/s)
+scored 2,048/3,676 choices (316.4/s)
+scored 2,560/3,676 choices (337.6/s)
+scored 3,072/3,676 choices (345.7/s)
+scored 3,584/3,676 choices (354.0/s)
+scored 3,676/3,676 choices (349.7/s)
+{
+  "benchmarks": {
+    "piqa": {
+      "acc": 0.6789989118607181,
+      "acc_norm": 0.6893362350380848,
+      "correct": 1248,
+      "correct_norm": 1267,
+      "elapsed_seconds": 10.618,
+      "examples": 1838
+    }
+  },
+  "chat_template_applied": false,
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_001994",
+  "checkpoint_step": 1994,
+  "custom_triton": true,
+  "dtype": "float16",
+  "scoring": "causal_choice_loglikelihood",
+  "sft_matched_eval_loss": null,
+  "zero_shot": true
+}
+{
+  "failures": {
+    "code_is_prime": [
+      "case_0_raised:NameError"
+    ],
+    "instruction_three_bullets": [
+      "long_bullets:0,1,2"
+    ],
+    "math_multiplication": [
+      "missing_number:391"
+    ]
+  },
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "passed": false,
+  "piqa_acc_norm": 0.6893362350380848,
+  "piqa_threshold": 0.6789
+}
+[evaluation] candidate failed promotion gate; report retained for comparison

reports/sft-v2-300k/evaluations/epoch_01_step_001994/piqa.json

@@ -0,0 +1,20 @@
+{
+  "benchmarks": {
+    "piqa": {
+      "acc": 0.6789989118607181,
+      "acc_norm": 0.6893362350380848,
+      "correct": 1248,
+      "correct_norm": 1267,
+      "elapsed_seconds": 10.618,
+      "examples": 1838
+    }
+  },
+  "chat_template_applied": false,
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_001994",
+  "checkpoint_step": 1994,
+  "custom_triton": true,
+  "dtype": "float16",
+  "scoring": "causal_choice_loglikelihood",
+  "sft_matched_eval_loss": null,
+  "zero_shot": true
+}

reports/sft-v2-300k/evaluations/epoch_01_step_001994/promotion.json

@@ -0,0 +1,17 @@
+{
+  "failures": {
+    "code_is_prime": [
+      "case_0_raised:NameError"
+    ],
+    "instruction_three_bullets": [
+      "long_bullets:0,1,2"
+    ],
+    "math_multiplication": [
+      "missing_number:391"
+    ]
+  },
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "passed": false,
+  "piqa_acc_norm": 0.6893362350380848,
+  "piqa_threshold": 0.6789
+}

reports/sft-v2-300k/evaluations/epoch_02_step_003988/chat.json

@@ -0,0 +1,130 @@
+{
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_003988",
+  "chat_template_applied": true,
+  "runtime": "pytorch",
+  "generation": {
+    "max_tokens": 512,
+    "temperature": 0.0,
+    "top_p": 1.0,
+    "top_k": 0,
+    "repetition_penalty": 1.1,
+    "repetition_context_size": 128
+  },
+  "results": [
+    {
+      "id": "greeting",
+      "prompt": "Hello!",
+      "response": "Hi there! How can I help you today?",
+      "repetition": {
+        "words": 8,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "math_multiplication",
+      "prompt": "Calculate 17 \u00d7 23. Give the result and one short calculation.",
+      "response": "To calculate 17 \u00d7 23, we can use the following steps:\n\nFirst, we need to find a number that is divisible by 17. This can be done by finding a multiple of 17 that divides into 17 evenly. For example, 17 \u00f7 17 = 1 remainder 0.\n\nNext, we need to find a number that is divisible by 23. This can be done by finding a multiple of 23 that divides into 23 evenly. For example, 23 \u00f7 23 = 1 remainder 2.\n\nNow, we can add these two numbers together to get 17 + 23 = 40.\n\nTherefore, 17 \u00d7 23 = 40.",
+      "repetition": {
+        "words": 95,
+        "distinct_2": 0.723404,
+        "repeated_trigrams": 18,
+        "repeated_trigram_ratio": 0.193548
+      }
+    },
+    {
+      "id": "math_word_problem",
+      "prompt": "A box contains 24 red balls and 18 blue balls. Seven balls are removed. How many remain? Answer with the number and one sentence.",
+      "response": "To solve this problem, we need to determine how many red balls are left after removing 7 balls from a box containing 24 red balls and 18 blue balls. \n\n1. **Understanding the Problem**: The problem states that there are 24 red balls and 18 blue balls in the box. We need to find out how many red balls are left after removing these balls.\n2. **Identifying the Required Information**: Since the boxes contain both red and blue balls, we can use the concept of \"remaining\" or \"unused.\" This means that if we remove one ball (the red ball), we have a remaining number of balls equal to the total number of balls minus 1.\n3. **Calculating Remaining Numbers**:\n   - If we remove one ball (the red ball), then the remaining number of balls is \\(24 - 1 = 17\\) balls.\n   - If we remove two balls (the blue ball), then the remaining number of balls is \\(18 - 2 = 16\\) balls.\n4. **Determining the Number of Red Balls**: Since there are 24 red balls and 18 blue balls, the total number of balls in the box is \\(24 + 18 = 44\\). Therefore, the number of red balls left after removing these balls is \\(44 - 1 = 43\\).\n5. **Calculating Remaining Numbers**:\n   - The number of red balls remains unchanged at 43.\n6. **Conclusion**: The problem states that there are 43 red balls and 17 blue balls in the box. This means that if we remove one ball (the red ball), we have a remaining number of balls equal to the total number of balls minus 1. Thus, the number of red balls left after removing these balls is \\(43 - 1 = 38\\).",
+      "repetition": {
+        "words": 273,
+        "distinct_2": 0.558824,
+        "repeated_trigrams": 92,
+        "repeated_trigram_ratio": 0.339483
+      }
+    },
+    {
+      "id": "code_add",
+      "prompt": "Write only executable Python code defining add(a, b), which returns the sum of a and b.",
+      "response": "Here's an example of how to define a function `add` that takes two arguments (a and b) and returns their sum:\n\n```python\ndef add(a, b):\n    return a + b\n```",
+      "repetition": {
+        "words": 29,
+        "distinct_2": 0.964286,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "code_is_prime",
+      "prompt": "Write only executable Python code defining is_prime(n). It must return True exactly when integer n is prime.",
+      "response": "Here's the complete implementation of the function `is_prime` that checks if a number is prime using the Sieve of Eratosthenes algorithm:\n\n```python\ndef is_prime(n):\n    if n <= 1:\n        return False\n    for i in range(2, int(n**0.5) + 1):\n        if n % i == 0:\n            return False\n    return True\n\ndef main():\n    n = int(input(\"Enter a number: \"))\n    print(is_prime(n))\n\nif __name__ == \"__main__\":\n    main()\n```",
+      "repetition": {
+        "words": 63,
+        "distinct_2": 0.919355,
+        "repeated_trigrams": 1,
+        "repeated_trigram_ratio": 0.016393
+      }
+    },
+    {
+      "id": "memory_name",
+      "prompt": [
+        {
+          "role": "user",
+          "content": "My name is Boris. Please remember it."
+        },
+        {
+          "role": "assistant",
+          "content": "Understood, your name is Boris."
+        },
+        {
+          "role": "user",
+          "content": "What is my name? Answer in one short sentence."
+        }
+      ],
+      "response": "My name is Boris.",
+      "repetition": {
+        "words": 4,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "memory_rules",
+      "prompt": [
+        {
+          "role": "user",
+          "content": "For our game, the secret piece is a blue triangle and collecting a star gives 3 points."
+        },
+        {
+          "role": "assistant",
+          "content": "Got it: blue triangle, and each star is worth 3 points."
+        },
+        {
+          "role": "user",
+          "content": "Remind me of the secret piece and the star score."
+        }
+      ],
+      "response": "The secret piece is a blue triangle, and each star is worth 3 points. The star score is the sum of all the stars in the game.",
+      "repetition": {
+        "words": 27,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "instruction_three_bullets",
+      "prompt": "Give exactly three bullet points for saving electricity. Use no more than eight words per bullet.",
+      "response": "* **Energy Efficiency**:  \n  - **Reduce Energy Consumption**:  \n    - Turn off lights and electronics when not in use, unplug devices when not in use, and use energy-efficient appliances.  \n    - Install smart thermostats or energy-harvesting systems to optimize heating/cooling and reduce energy waste.",
+      "repetition": {
+        "words": 40,
+        "distinct_2": 0.897436,
+        "repeated_trigrams": 2,
+        "repeated_trigram_ratio": 0.052632
+      }
+    }
+  ]
+}

reports/sft-v2-300k/evaluations/epoch_02_step_003988/evaluation.log

@@ -0,0 +1,123 @@
+Activated conda/uv virtual environment at /venv/main
+[Tokenizer] Loaded from /workspace/tr-hash-moe-200m-sft-v2-300k/tokenized/tr-hash-32k-v2-2048/tokenizer
+
+[greeting]
+Hi there! How can I help you today?
+
+[math_multiplication]
+To calculate 17 × 23, we can use the following steps:
+
+First, we need to find a number that is divisible by 17. This can be done by finding a multiple of 17 that divides into 17 evenly. For example, 17 ÷ 17 = 1 remainder 0.
+
+Next, we need to find a number that is divisible by 23. This can be done by finding a multiple of 23 that divides into 23 evenly. For example, 23 ÷ 23 = 1 remainder 2.
+
+Now, we can add these two numbers together to get 17 + 23 = 40.
+
+Therefore, 17 × 23 = 40.
+
+[math_word_problem]
+To solve this problem, we need to determine how many red balls are left after removing 7 balls from a box containing 24 red balls and 18 blue balls. 
+
+1. **Understanding the Problem**: The problem states that there are 24 red balls and 18 blue balls in the box. We need to find out how many red balls are left after removing these balls.
+2. **Identifying the Required Information**: Since the boxes contain both red and blue balls, we can use the concept of "remaining" or "unused." This means that if we remove one ball (the red ball), we have a remaining number of balls equal to the total number of balls minus 1.
+3. **Calculating Remaining Numbers**:
+   - If we remove one ball (the red ball), then the remaining number of balls is \(24 - 1 = 17\) balls.
+   - If we remove two balls (the blue ball), then the remaining number of balls is \(18 - 2 = 16\) balls.
+4. **Determining the Number of Red Balls**: Since there are 24 red balls and 18 blue balls, the total number of balls in the box is \(24 + 18 = 44\). Therefore, the number of red balls left after removing these balls is \(44 - 1 = 43\).
+5. **Calculating Remaining Numbers**:
+   - The number of red balls remains unchanged at 43.
+6. **Conclusion**: The problem states that there are 43 red balls and 17 blue balls in the box. This means that if we remove one ball (the red ball), we have a remaining number of balls equal to the total number of balls minus 1. Thus, the number of red balls left after removing these balls is \(43 - 1 = 38\).
+
+[code_add]
+Here's an example of how to define a function `add` that takes two arguments (a and b) and returns their sum:
+
+```python
+def add(a, b):
+    return a + b
+```
+
+[code_is_prime]
+Here's the complete implementation of the function `is_prime` that checks if a number is prime using the Sieve of Eratosthenes algorithm:
+
+```python
+def is_prime(n):
+    if n <= 1:
+        return False
+    for i in range(2, int(n**0.5) + 1):
+        if n % i == 0:
+            return False
+    return True
+
+def main():
+    n = int(input("Enter a number: "))
+    print(is_prime(n))
+
+if __name__ == "__main__":
+    main()
+```
+
+[memory_name]
+My name is Boris.
+
+[memory_rules]
+The secret piece is a blue triangle, and each star is worth 3 points. The star score is the sum of all the stars in the game.
+
+[instruction_three_bullets]
+* **Energy Efficiency**:  
+  - **Reduce Energy Consumption**:  
+    - Turn off lights and electronics when not in use, unplug devices when not in use, and use energy-efficient appliances.  
+    - Install smart thermostats or energy-harvesting systems to optimize heating/cooling and reduce energy waste.
+
+Report saved: artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_02_step_003988/chat.json
+22:13:54 | INFO | [device] backend=cuda device=cuda name=NVIDIA GeForce RTX 5090 matmul=cuBLAS/cuBLASLt distributed=NCCL sdpa=true flash_attention=true custom_triton=true
+[Tokenizer] Loaded from /workspace/tr-hash-moe-200m-sft-v2-300k/tokenized/tr-hash-32k-v2-2048/tokenizer
+scored 512/3,676 choices (213.4/s)
+scored 1,024/3,676 choices (410.9/s)
+scored 1,536/3,676 choices (594.3/s)
+scored 2,048/3,676 choices (632.6/s)
+scored 2,560/3,676 choices (589.0/s)
+scored 3,072/3,676 choices (544.1/s)
+scored 3,584/3,676 choices (520.2/s)
+scored 3,676/3,676 choices (505.5/s)
+{
+  "benchmarks": {
+    "piqa": {
+      "acc": 0.6784548422198041,
+      "acc_norm": 0.6882480957562568,
+      "correct": 1247,
+      "correct_norm": 1265,
+      "elapsed_seconds": 7.368,
+      "examples": 1838
+    }
+  },
+  "chat_template_applied": false,
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_003988",
+  "checkpoint_step": 3988,
+  "custom_triton": true,
+  "dtype": "float16",
+  "scoring": "causal_choice_loglikelihood",
+  "sft_matched_eval_loss": null,
+  "zero_shot": true
+}
+{
+  "failures": {
+    "code_is_prime": [
+      "python_exec_error:NameError"
+    ],
+    "instruction_three_bullets": [
+      "bullet_count:4",
+      "long_bullets:2,3"
+    ],
+    "math_multiplication": [
+      "missing_number:391"
+    ],
+    "math_word_problem": [
+      "missing_number:35"
+    ]
+  },
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "passed": false,
+  "piqa_acc_norm": 0.6882480957562568,
+  "piqa_threshold": 0.6789
+}
+[evaluation] candidate failed promotion gate; report retained for comparison

reports/sft-v2-300k/evaluations/epoch_02_step_003988/piqa.json

@@ -0,0 +1,20 @@
+{
+  "benchmarks": {
+    "piqa": {
+      "acc": 0.6784548422198041,
+      "acc_norm": 0.6882480957562568,
+      "correct": 1247,
+      "correct_norm": 1265,
+      "elapsed_seconds": 7.368,
+      "examples": 1838
+    }
+  },
+  "chat_template_applied": false,
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_003988",
+  "checkpoint_step": 3988,
+  "custom_triton": true,
+  "dtype": "float16",
+  "scoring": "causal_choice_loglikelihood",
+  "sft_matched_eval_loss": null,
+  "zero_shot": true
+}

reports/sft-v2-300k/evaluations/epoch_02_step_003988/promotion.json

@@ -0,0 +1,21 @@
+{
+  "failures": {
+    "code_is_prime": [
+      "python_exec_error:NameError"
+    ],
+    "instruction_three_bullets": [
+      "bullet_count:4",
+      "long_bullets:2,3"
+    ],
+    "math_multiplication": [
+      "missing_number:391"
+    ],
+    "math_word_problem": [
+      "missing_number:35"
+    ]
+  },
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "passed": false,
+  "piqa_acc_norm": 0.6882480957562568,
+  "piqa_threshold": 0.6789
+}

reports/sft-v2-300k/evaluations/epoch_03_step_005982/chat.json

@@ -0,0 +1,130 @@
+{
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+  "chat_template_applied": true,
+  "runtime": "pytorch",
+  "generation": {
+    "max_tokens": 512,
+    "temperature": 0.0,
+    "top_p": 1.0,
+    "top_k": 0,
+    "repetition_penalty": 1.1,
+    "repetition_context_size": 128
+  },
+  "results": [
+    {
+      "id": "greeting",
+      "prompt": "Hello!",
+      "response": "Hi there! How can I help you today?",
+      "repetition": {
+        "words": 8,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "math_multiplication",
+      "prompt": "Calculate 17 \u00d7 23. Give the result and one short calculation.",
+      "response": "To calculate 17 \u00d7 23, we can use the following steps:\n\nFirst, we need to find a number that is divisible by 17. This can be done by finding a multiple of 17 less than or equal to 23. We can do this by dividing 23 by 17 and taking the remainder.\n\nNext, we need to find a number that is divisible by 23. This can be done by finding a multiple of 23 less than or equal to 17. We can do this by dividing 17 by 23 and taking the remainder.\n\nFinally, we add up all the results from both steps to get the final answer.\n\nTherefore, the result of 17 \u00d7 23 is $\\boxed{17}$.",
+      "repetition": {
+        "words": 116,
+        "distinct_2": 0.695652,
+        "repeated_trigrams": 25,
+        "repeated_trigram_ratio": 0.219298
+      }
+    },
+    {
+      "id": "math_word_problem",
+      "prompt": "A box contains 24 red balls and 18 blue balls. Seven balls are removed. How many remain? Answer with the number and one sentence.",
+      "response": "The total number of balls is \\(24 + 18 = 42\\). Remaining balls are \\(42 - 7 = 35\\) balls. \n\n**Answer:** The remaining balls are 35.",
+      "repetition": {
+        "words": 22,
+        "distinct_2": 0.904762,
+        "repeated_trigrams": 1,
+        "repeated_trigram_ratio": 0.05
+      }
+    },
+    {
+      "id": "code_add",
+      "prompt": "Write only executable Python code defining add(a, b), which returns the sum of a and b.",
+      "response": "Here's an example of how to define a function `add` that adds two numbers:\n\n```python\ndef add(a, b):\n    return a + b\n```",
+      "repetition": {
+        "words": 22,
+        "distinct_2": 0.952381,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "code_is_prime",
+      "prompt": "Write only executable Python code defining is_prime(n). It must return True exactly when integer n is prime.",
+      "response": "Here's the complete implementation of the function `is_prime` that checks if a number is prime using the Sieve of Eratosthenes algorithm:\n\n```python\ndef is_prime(n):\n    if n < 2:\n        return False\n    for i in range(2, int(n**0.5) + 1):\n        if n % i == 0:\n            return False\n    return True\n\ndef main():\n    n = int(input(\"Enter a positive integer: \"))\n    print(is_prime(n))\n\nif __name__ == \"__main__\":\n    main()\n```",
+      "repetition": {
+        "words": 64,
+        "distinct_2": 0.936508,
+        "repeated_trigrams": 1,
+        "repeated_trigram_ratio": 0.016129
+      }
+    },
+    {
+      "id": "memory_name",
+      "prompt": [
+        {
+          "role": "user",
+          "content": "My name is Boris. Please remember it."
+        },
+        {
+          "role": "assistant",
+          "content": "Understood, your name is Boris."
+        },
+        {
+          "role": "user",
+          "content": "What is my name? Answer in one short sentence."
+        }
+      ],
+      "response": "My name is Boris.",
+      "repetition": {
+        "words": 4,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "memory_rules",
+      "prompt": [
+        {
+          "role": "user",
+          "content": "For our game, the secret piece is a blue triangle and collecting a star gives 3 points."
+        },
+        {
+          "role": "assistant",
+          "content": "Got it: blue triangle, and each star is worth 3 points."
+        },
+        {
+          "role": "user",
+          "content": "Remind me of the secret piece and the star score."
+        }
+      ],
+      "response": "The secret piece is a blue triangle with a star on its top. Each star is worth 3 points.",
+      "repetition": {
+        "words": 19,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    },
+    {
+      "id": "instruction_three_bullets",
+      "prompt": "Give exactly three bullet points for saving electricity. Use no more than eight words per bullet.",
+      "response": "* Turn off lights and electronics when not in use to reduce energy consumption.\n* Use energy-efficient appliances and LED lighting to lower utility bills.\n* Install solar panels or other renewable energy sources to generate power during non-peak hours.",
+      "repetition": {
+        "words": 39,
+        "distinct_2": 1.0,
+        "repeated_trigrams": 0,
+        "repeated_trigram_ratio": 0.0
+      }
+    }
+  ]
+}

reports/sft-v2-300k/evaluations/epoch_03_step_005982/evaluation.log

@@ -0,0 +1,110 @@
+Activated conda/uv virtual environment at /venv/main
+[Tokenizer] Loaded from /workspace/tr-hash-moe-200m-sft-v2-300k/tokenized/tr-hash-32k-v2-2048/tokenizer
+
+[greeting]
+Hi there! How can I help you today?
+
+[math_multiplication]
+To calculate 17 × 23, we can use the following steps:
+
+First, we need to find a number that is divisible by 17. This can be done by finding a multiple of 17 less than or equal to 23. We can do this by dividing 23 by 17 and taking the remainder.
+
+Next, we need to find a number that is divisible by 23. This can be done by finding a multiple of 23 less than or equal to 17. We can do this by dividing 17 by 23 and taking the remainder.
+
+Finally, we add up all the results from both steps to get the final answer.
+
+Therefore, the result of 17 × 23 is $\boxed{17}$.
+
+[math_word_problem]
+The total number of balls is \(24 + 18 = 42\). Remaining balls are \(42 - 7 = 35\) balls. 
+
+**Answer:** The remaining balls are 35.
+
+[code_add]
+Here's an example of how to define a function `add` that adds two numbers:
+
+```python
+def add(a, b):
+    return a + b
+```
+
+[code_is_prime]
+Here's the complete implementation of the function `is_prime` that checks if a number is prime using the Sieve of Eratosthenes algorithm:
+
+```python
+def is_prime(n):
+    if n < 2:
+        return False
+    for i in range(2, int(n**0.5) + 1):
+        if n % i == 0:
+            return False
+    return True
+
+def main():
+    n = int(input("Enter a positive integer: "))
+    print(is_prime(n))
+
+if __name__ == "__main__":
+    main()
+```
+
+[memory_name]
+My name is Boris.
+
+[memory_rules]
+The secret piece is a blue triangle with a star on its top. Each star is worth 3 points.
+
+[instruction_three_bullets]
+* Turn off lights and electronics when not in use to reduce energy consumption.
+* Use energy-efficient appliances and LED lighting to lower utility bills.
+* Install solar panels or other renewable energy sources to generate power during non-peak hours.
+
+Report saved: artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_03_step_005982/chat.json
+22:13:49 | INFO | [device] backend=cuda device=cuda name=NVIDIA GeForce RTX 5090 matmul=cuBLAS/cuBLASLt distributed=NCCL sdpa=true flash_attention=true custom_triton=true
+[Tokenizer] Loaded from /workspace/tr-hash-moe-200m-sft-v2-300k/tokenized/tr-hash-32k-v2-2048/tokenizer
+scored 512/3,676 choices (97.0/s)
+scored 1,024/3,676 choices (170.7/s)
+scored 1,536/3,676 choices (233.9/s)
+scored 2,048/3,676 choices (276.1/s)
+scored 2,560/3,676 choices (300.3/s)
+scored 3,072/3,676 choices (312.7/s)
+scored 3,584/3,676 choices (323.8/s)
+scored 3,676/3,676 choices (321.1/s)
+{
+  "benchmarks": {
+    "piqa": {
+      "acc": 0.6800870511425462,
+      "acc_norm": 0.690968443960827,
+      "correct": 1250,
+      "correct_norm": 1270,
+      "elapsed_seconds": 11.556,
+      "examples": 1838
+    }
+  },
+  "chat_template_applied": false,
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+  "checkpoint_step": 5982,
+  "custom_triton": true,
+  "dtype": "float16",
+  "scoring": "causal_choice_loglikelihood",
+  "sft_matched_eval_loss": null,
+  "zero_shot": true
+}
+{
+  "failures": {
+    "code_is_prime": [
+      "python_exec_error:NameError"
+    ],
+    "instruction_three_bullets": [
+      "long_bullets:0,1,2"
+    ],
+    "math_multiplication": [
+      "missing_number:391"
+    ]
+  },
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "passed": false,
+  "piqa_acc_norm": 0.690968443960827,
+  "piqa_threshold": 0.6789
+}
+[evaluation] candidate failed promotion gate; report retained for comparison

reports/sft-v2-300k/evaluations/epoch_03_step_005982/piqa.json

@@ -0,0 +1,20 @@
+{
+  "benchmarks": {
+    "piqa": {
+      "acc": 0.6800870511425462,
+      "acc_norm": 0.690968443960827,
+      "correct": 1250,
+      "correct_norm": 1270,
+      "elapsed_seconds": 11.556,
+      "examples": 1838
+    }
+  },
+  "chat_template_applied": false,
+  "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+  "checkpoint_step": 5982,
+  "custom_triton": true,
+  "dtype": "float16",
+  "scoring": "causal_choice_loglikelihood",
+  "sft_matched_eval_loss": null,
+  "zero_shot": true
+}

reports/sft-v2-300k/evaluations/epoch_03_step_005982/promotion.json

@@ -0,0 +1,17 @@
+{
+  "failures": {
+    "code_is_prime": [
+      "python_exec_error:NameError"
+    ],
+    "instruction_three_bullets": [
+      "long_bullets:0,1,2"
+    ],
+    "math_multiplication": [
+      "missing_number:391"
+    ]
+  },
+  "panel_id": "tr-hash-moe-200m-clean-sft-v2-regression",
+  "passed": false,
+  "piqa_acc_norm": 0.690968443960827,
+  "piqa_threshold": 0.6789
+}

reports/sft-v2-300k/evaluations/summary.json

@@ -0,0 +1,84 @@
+{
+  "candidates": [
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_001994",
+      "epoch": 1,
+      "matched_eval_loss": 0.990943,
+      "matched_eval_ppl": 2.69,
+      "piqa_acc": 0.6789989118607181,
+      "piqa_acc_norm": 0.6893362350380848,
+      "piqa_correct": 1248,
+      "piqa_correct_norm": 1267,
+      "promotion_failures": {
+        "code_is_prime": [
+          "case_0_raised:NameError"
+        ],
+        "instruction_three_bullets": [
+          "long_bullets:0,1,2"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_01_step_001994",
+      "step": 1994
+    },
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_003988",
+      "epoch": 2,
+      "matched_eval_loss": 0.963912,
+      "matched_eval_ppl": 2.62,
+      "piqa_acc": 0.6784548422198041,
+      "piqa_acc_norm": 0.6882480957562568,
+      "piqa_correct": 1247,
+      "piqa_correct_norm": 1265,
+      "promotion_failures": {
+        "code_is_prime": [
+          "python_exec_error:NameError"
+        ],
+        "instruction_three_bullets": [
+          "bullet_count:4",
+          "long_bullets:2,3"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ],
+        "math_word_problem": [
+          "missing_number:35"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_02_step_003988",
+      "step": 3988
+    },
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+      "epoch": 3,
+      "matched_eval_loss": 0.959617,
+      "matched_eval_ppl": 2.61,
+      "piqa_acc": 0.6800870511425462,
+      "piqa_acc_norm": 0.690968443960827,
+      "piqa_correct": 1250,
+      "piqa_correct_norm": 1270,
+      "promotion_failures": {
+        "code_is_prime": [
+          "python_exec_error:NameError"
+        ],
+        "instruction_three_bullets": [
+          "long_bullets:0,1,2"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_03_step_005982",
+      "step": 5982
+    }
+  ],
+  "release_ready": false,
+  "schema_version": 1,
+  "selected": null,
+  "selection_policy": "promotion gate pass, then maximum PIQA acc_norm, maximum PIQA acc, minimum matched SFT eval loss, earliest epoch"
+}

reports/sft-v2-300k/evaluations/summary_epoch3_release.json

@@ -0,0 +1,107 @@
+{
+  "candidates": [
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_001994",
+      "epoch": 1,
+      "matched_eval_loss": 0.990943,
+      "matched_eval_ppl": 2.69,
+      "piqa_acc": 0.6789989118607181,
+      "piqa_acc_norm": 0.6893362350380848,
+      "piqa_correct": 1248,
+      "piqa_correct_norm": 1267,
+      "promotion_failures": {
+        "code_is_prime": [
+          "case_0_raised:NameError"
+        ],
+        "instruction_three_bullets": [
+          "long_bullets:0,1,2"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_01_step_001994",
+      "step": 1994
+    },
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_003988",
+      "epoch": 2,
+      "matched_eval_loss": 0.963912,
+      "matched_eval_ppl": 2.62,
+      "piqa_acc": 0.6784548422198041,
+      "piqa_acc_norm": 0.6882480957562568,
+      "piqa_correct": 1247,
+      "piqa_correct_norm": 1265,
+      "promotion_failures": {
+        "code_is_prime": [
+          "python_exec_error:NameError"
+        ],
+        "instruction_three_bullets": [
+          "bullet_count:4",
+          "long_bullets:2,3"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ],
+        "math_word_problem": [
+          "missing_number:35"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_02_step_003988",
+      "step": 3988
+    },
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+      "epoch": 3,
+      "matched_eval_loss": 0.959617,
+      "matched_eval_ppl": 2.61,
+      "piqa_acc": 0.6800870511425462,
+      "piqa_acc_norm": 0.690968443960827,
+      "piqa_correct": 1250,
+      "piqa_correct_norm": 1270,
+      "promotion_failures": {
+        "code_is_prime": [
+          "python_exec_error:NameError"
+        ],
+        "instruction_three_bullets": [
+          "long_bullets:0,1,2"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_03_step_005982",
+      "step": 5982
+    }
+  ],
+  "release_ready": true,
+  "schema_version": 1,
+  "selected": {
+    "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+    "epoch": 3,
+    "matched_eval_loss": 0.959617,
+    "matched_eval_ppl": 2.61,
+    "piqa_acc": 0.6800870511425462,
+    "piqa_acc_norm": 0.690968443960827,
+    "piqa_correct": 1250,
+    "piqa_correct_norm": 1270,
+    "promotion_failures": {
+      "code_is_prime": [
+        "python_exec_error:NameError"
+      ],
+      "instruction_three_bullets": [
+        "long_bullets:0,1,2"
+      ],
+      "math_multiplication": [
+        "missing_number:391"
+      ]
+    },
+    "promotion_passed": false,
+    "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_03_step_005982",
+    "step": 5982
+  },
+  "selection_policy": "manual promotion of epoch 3: highest full-PIQA acc_norm and acc, lowest held-out SFT loss; strict behavior failures retained"
+}

reports/sft-v2-300k/metrics.csv

@@ -0,0 +1,604 @@
+step,train_loss,train_ppl,matched_eval_loss,matched_eval_ppl,lr,expert_lr,tok_s,supervised_tokens,min_label,max_label,bad_labels,matched_eval_tokens,natural_eval_loss,natural_eval_ppl,natural_eval_tokens
+0,,,1.722610,5.60,1.117318e-07,,,,,,0,981892,,,0
+1,1.784332,5.96,,,2.234637e-07,2.234637e-07,33004,12740,0,31862,0,,,,
+10,1.580614,4.86,,,1.229050e-06,1.229050e-06,303685,11714,0,31758,0,,,,
+20,1.595486,4.93,,,2.346369e-06,2.346369e-06,311939,11868,0,31890,0,,,,
+30,1.417899,4.13,,,3.463687e-06,3.463687e-06,311824,11754,0,31942,0,,,,
+40,1.268079,3.55,,,4.581006e-06,4.581006e-06,311511,12082,0,31909,0,,,,
+50,1.488278,4.43,,,5.698324e-06,5.698324e-06,311085,12621,0,31862,0,,,,
+60,1.362671,3.91,,,6.815642e-06,6.815642e-06,294456,9639,0,31996,0,,,,
+70,1.251097,3.49,,,7.932961e-06,7.932961e-06,310533,14122,0,31890,0,,,,
+80,1.185581,3.27,,,9.050279e-06,9.050279e-06,310171,9267,0,31963,0,,,,
+90,1.271076,3.56,,,1.016760e-05,1.016760e-05,310166,12705,0,31862,0,,,,
+100,1.224002,3.40,,,1.128492e-05,1.128492e-05,309672,11751,0,31862,0,,,,
+110,1.146304,3.15,,,1.240223e-05,1.240223e-05,292943,10691,0,31925,0,,,,
+120,1.250084,3.49,,,1.351955e-05,1.351955e-05,309374,12023,0,31882,0,,,,
+130,1.178241,3.25,,,1.463687e-05,1.463687e-05,309043,11593,0,31847,0,,,,
+140,1.139647,3.13,,,1.575419e-05,1.575419e-05,308913,9488,0,31865,0,,,,
+150,1.017182,2.77,,,1.687151e-05,1.687151e-05,308762,13434,0,31758,0,,,,
+160,1.180706,3.26,,,1.798883e-05,1.798883e-05,291768,11782,0,31897,0,,,,
+170,1.048072,2.85,,,1.910615e-05,1.910615e-05,308464,13066,0,31810,0,,,,
+180,1.216714,3.38,,,1.999999e-05,1.999999e-05,308090,9115,0,31996,0,,,,
+190,1.081634,2.95,,,1.999981e-05,1.999981e-05,307976,11841,0,31854,0,,,,
+200,1.133811,3.11,,,1.999936e-05,1.999936e-05,307805,10843,0,31862,0,,,,
+210,1.148897,3.15,,,1.999865e-05,1.999865e-05,291650,7773,0,31923,0,,,,
+220,1.119341,3.06,,,1.999767e-05,1.999767e-05,307559,10632,0,31862,0,,,,
+230,1.198862,3.32,,,1.999643e-05,1.999643e-05,307712,11939,0,31972,0,,,,
+240,1.106810,3.02,,,1.999493e-05,1.999493e-05,307192,11537,0,31930,0,,,,
+250,1.145610,3.14,,,1.999316e-05,1.999316e-05,307503,13450,0,31951,0,,,,
+260,1.066673,2.91,,,1.999113e-05,1.999113e-05,291379,15182,0,31862,0,,,,
+270,1.286960,3.62,,,1.998884e-05,1.998884e-05,307212,11153,0,31992,0,,,,
+280,1.136745,3.12,,,1.998628e-05,1.998628e-05,307309,11061,0,31700,0,,,,
+290,1.106904,3.02,,,1.998346e-05,1.998346e-05,307114,15886,0,31862,0,,,,
+300,0.984750,2.68,,,1.998038e-05,1.998038e-05,307128,9236,0,31996,0,,,,
+310,1.119017,3.06,,,1.997703e-05,1.997703e-05,291245,12671,0,31993,0,,,,
+320,1.177243,3.25,,,1.997342e-05,1.997342e-05,306855,9272,0,31893,0,,,,
+330,1.104771,3.02,,,1.996955e-05,1.996955e-05,306866,10141,0,31870,0,,,,
+340,0.987604,2.68,,,1.996541e-05,1.996541e-05,306680,13082,0,31817,0,,,,
+350,1.124717,3.08,,,1.996101e-05,1.996101e-05,306794,12790,0,31865,0,,,,
+360,1.017624,2.77,,,1.995635e-05,1.995635e-05,290982,10980,0,31758,0,,,,
+370,1.053635,2.87,,,1.995142e-05,1.995142e-05,306560,13711,0,31862,0,,,,
+380,0.991393,2.69,,,1.994624e-05,1.994624e-05,306492,11244,0,31758,0,,,,
+390,1.050099,2.86,,,1.994079e-05,1.994079e-05,306485,12426,0,31895,0,,,,
+400,1.146371,3.15,,,1.993508e-05,1.993508e-05,306663,12000,0,31918,0,,,,
+410,1.168288,3.22,,,1.992911e-05,1.992911e-05,290635,13419,0,31980,0,,,,
+420,1.074304,2.93,,,1.992287e-05,1.992287e-05,306308,11196,0,31758,0,,,,
+430,0.958106,2.61,,,1.991638e-05,1.991638e-05,306174,15005,0,31996,0,,,,
+440,1.141989,3.13,,,1.990962e-05,1.990962e-05,306293,9943,0,31963,0,,,,
+450,0.996209,2.71,,,1.990260e-05,1.990260e-05,306170,14077,0,31758,0,,,,
+460,1.005229,2.73,,,1.989532e-05,1.989532e-05,290663,11374,0,31963,0,,,,
+470,1.244940,3.47,,,1.988778e-05,1.988778e-05,306246,11398,0,31912,0,,,,
+480,0.962253,2.62,,,1.987998e-05,1.987998e-05,306141,11679,0,31996,0,,,,
+490,1.054399,2.87,,,1.987192e-05,1.987192e-05,306232,11773,0,31849,0,,,,
+500,1.199414,3.32,,,1.986360e-05,1.986360e-05,306249,9583,0,31794,0,,,,
+510,1.071409,2.92,,,1.985502e-05,1.985502e-05,290722,10538,0,31992,0,,,,
+520,1.042976,2.84,,,1.984618e-05,1.984618e-05,306329,12555,0,31963,0,,,,
+530,1.008838,2.74,,,1.983708e-05,1.983708e-05,306022,12458,0,31996,0,,,,
+540,1.083239,2.95,,,1.982772e-05,1.982772e-05,306272,13860,0,31857,0,,,,
+550,1.141052,3.13,,,1.981810e-05,1.981810e-05,305941,15059,0,31918,0,,,,
+560,0.962177,2.62,,,1.980823e-05,1.980823e-05,289195,10933,0,31771,0,,,,
+570,0.983692,2.67,,,1.979809e-05,1.979809e-05,306057,13375,0,31993,0,,,,
+580,1.160142,3.19,,,1.978770e-05,1.978770e-05,306076,8184,0,31826,0,,,,
+590,1.007020,2.74,,,1.977705e-05,1.977705e-05,305933,11159,0,31916,0,,,,
+600,1.089630,2.97,,,1.976615e-05,1.976615e-05,306095,12960,0,31862,0,,,,
+610,1.038440,2.82,,,1.975498e-05,1.975498e-05,290202,7994,0,31902,0,,,,
+620,1.031688,2.81,,,1.974356e-05,1.974356e-05,305954,12927,0,31930,0,,,,
+630,1.046838,2.85,,,1.973189e-05,1.973189e-05,306124,10395,0,31937,0,,,,
+640,1.028803,2.80,,,1.971996e-05,1.971996e-05,305916,9367,0,31749,0,,,,
+650,1.053077,2.87,,,1.970777e-05,1.970777e-05,305954,11661,0,31963,0,,,,
+660,1.010500,2.75,,,1.969533e-05,1.969533e-05,290296,12491,0,31862,0,,,,
+670,1.031570,2.81,,,1.968263e-05,1.968263e-05,306091,12053,0,31862,0,,,,
+680,1.070909,2.92,,,1.966968e-05,1.966968e-05,305849,12392,0,31890,0,,,,
+690,1.101183,3.01,,,1.965647e-05,1.965647e-05,305948,11247,0,31914,0,,,,
+700,1.065088,2.90,,,1.964301e-05,1.964301e-05,306057,11353,0,31963,0,,,,
+710,0.997009,2.71,,,1.962930e-05,1.962930e-05,290122,11911,0,31990,0,,,,
+720,1.052592,2.87,,,1.961533e-05,1.961533e-05,305907,11470,0,31996,0,,,,
+730,1.061573,2.89,,,1.960111e-05,1.960111e-05,306019,15000,0,31862,0,,,,
+740,1.054696,2.87,,,1.958664e-05,1.958664e-05,306058,13861,0,31990,0,,,,
+750,0.996264,2.71,,,1.957192e-05,1.957192e-05,305856,12429,0,31993,0,,,,
+760,1.011571,2.75,,,1.955695e-05,1.955695e-05,289293,12034,0,31993,0,,,,
+770,1.032849,2.81,,,1.954172e-05,1.954172e-05,305923,14298,0,31849,0,,,,
+780,1.023882,2.78,,,1.952625e-05,1.952625e-05,305796,12099,0,31930,0,,,,
+790,0.942202,2.57,,,1.951052e-05,1.951052e-05,305805,13839,0,31993,0,,,,
+800,1.065145,2.90,,,1.949455e-05,1.949455e-05,306040,13714,0,31890,0,,,,
+810,0.890659,2.44,,,1.947832e-05,1.947832e-05,290299,13167,0,31993,0,,,,
+820,0.946069,2.58,,,1.946185e-05,1.946185e-05,305748,16087,0,31951,0,,,,
+830,1.021013,2.78,,,1.944513e-05,1.944513e-05,305814,12592,0,31993,0,,,,
+840,0.945713,2.57,,,1.942817e-05,1.942817e-05,306079,11756,0,31699,0,,,,
+850,0.985645,2.68,,,1.941095e-05,1.941095e-05,305747,10704,0,31758,0,,,,
+860,1.025467,2.79,,,1.939349e-05,1.939349e-05,289829,12185,0,31963,0,,,,
+870,1.023195,2.78,,,1.937579e-05,1.937579e-05,305946,11608,0,31918,0,,,,
+880,0.971718,2.64,,,1.935783e-05,1.935783e-05,305656,12334,0,31996,0,,,,
+890,0.907854,2.48,,,1.933964e-05,1.933964e-05,305774,11144,0,31989,0,,,,
+900,1.012065,2.75,,,1.932119e-05,1.932119e-05,306010,12140,0,31847,0,,,,
+910,1.115264,3.05,,,1.930251e-05,1.930251e-05,289768,12253,0,31882,0,,,,
+920,0.891556,2.44,,,1.928358e-05,1.928358e-05,305886,10922,0,31993,0,,,,
+930,1.076823,2.94,,,1.926441e-05,1.926441e-05,305820,9621,0,31980,0,,,,
+940,1.053237,2.87,,,1.924500e-05,1.924500e-05,305930,11090,0,31924,0,,,,
+950,1.144381,3.14,,,1.922534e-05,1.922534e-05,305931,13867,0,31913,0,,,,
+960,0.889150,2.43,,,1.920544e-05,1.920544e-05,289868,13823,0,31890,0,,,,
+970,1.036519,2.82,,,1.918531e-05,1.918531e-05,305719,11798,0,31996,0,,,,
+980,1.114385,3.05,,,1.916493e-05,1.916493e-05,305563,12001,0,31982,0,,,,
+990,0.998184,2.71,,,1.914431e-05,1.914431e-05,305591,10980,0,31890,0,,,,
+1000,1.019749,2.77,,,1.912346e-05,1.912346e-05,305473,10936,0,31953,0,,,,
+1010,1.113383,3.04,,,1.910237e-05,1.910237e-05,289408,11832,0,31702,0,,,,
+1020,0.919322,2.51,,,1.908104e-05,1.908104e-05,305862,10785,0,31742,0,,,,
+1030,1.011782,2.75,,,1.905947e-05,1.905947e-05,305808,12059,0,31890,0,,,,
+1040,1.005398,2.73,,,1.903767e-05,1.903767e-05,306013,10685,0,31996,0,,,,
+1050,1.062743,2.89,,,1.901563e-05,1.901563e-05,305656,12637,0,31992,0,,,,
+1060,1.037020,2.82,,,1.899335e-05,1.899335e-05,288436,14087,0,31993,0,,,,
+1070,1.002725,2.73,,,1.897085e-05,1.897085e-05,296077,13762,0,31858,0,,,,
+1080,1.077865,2.94,,,1.894810e-05,1.894810e-05,306122,8027,0,31996,0,,,,
+1090,1.182995,3.26,,,1.892513e-05,1.892513e-05,305314,13503,0,31921,0,,,,
+1100,0.996084,2.71,,,1.890192e-05,1.890192e-05,305553,10803,0,31758,0,,,,
+1110,0.901468,2.46,,,1.887848e-05,1.887848e-05,289306,12844,0,31807,0,,,,
+1120,0.932851,2.54,,,1.885481e-05,1.885481e-05,305650,12982,0,31982,0,,,,
+1130,1.061891,2.89,,,1.883091e-05,1.883091e-05,305646,13555,0,31969,0,,,,
+1140,1.062364,2.89,,,1.880679e-05,1.880679e-05,305513,10376,0,31998,0,,,,
+1150,0.985444,2.68,,,1.878243e-05,1.878243e-05,305624,11935,0,31993,0,,,,
+1160,0.981351,2.67,,,1.875784e-05,1.875784e-05,289617,13056,0,31862,0,,,,
+1170,0.929571,2.53,,,1.873303e-05,1.873303e-05,305728,10215,0,31890,0,,,,
+1180,0.977079,2.66,,,1.870799e-05,1.870799e-05,305808,12478,0,31993,0,,,,
+1190,1.029162,2.80,,,1.868272e-05,1.868272e-05,305469,12097,0,31907,0,,,,
+1200,1.042365,2.84,,,1.865723e-05,1.865723e-05,305491,9568,0,31884,0,,,,
+1210,1.016318,2.76,,,1.863151e-05,1.863151e-05,289865,13878,0,31972,0,,,,
+1220,1.052991,2.87,,,1.860558e-05,1.860558e-05,305654,15360,0,31993,0,,,,
+1230,1.040995,2.83,,,1.857941e-05,1.857941e-05,305626,12871,0,31917,0,,,,
+1240,1.032075,2.81,,,1.855303e-05,1.855303e-05,305639,10535,0,31778,0,,,,
+1250,1.036100,2.82,,,1.852642e-05,1.852642e-05,305821,12329,0,31996,0,,,,
+1260,0.983745,2.67,,,1.849960e-05,1.849960e-05,289072,13261,0,31963,0,,,,
+1270,1.043202,2.84,,,1.847255e-05,1.847255e-05,305522,12216,0,31993,0,,,,
+1280,0.984206,2.68,,,1.844529e-05,1.844529e-05,305445,13456,0,31996,0,,,,
+1290,1.035801,2.82,,,1.841780e-05,1.841780e-05,305569,12496,0,31977,0,,,,
+1300,0.940103,2.56,,,1.839010e-05,1.839010e-05,305773,13683,0,31976,0,,,,
+1310,0.921288,2.51,,,1.836218e-05,1.836218e-05,290069,13039,0,31947,0,,,,
+1320,1.060234,2.89,,,1.833405e-05,1.833405e-05,305604,14121,0,31998,0,,,,
+1330,0.915440,2.50,,,1.830570e-05,1.830570e-05,305448,10040,0,31970,0,,,,
+1340,1.044218,2.84,,,1.827714e-05,1.827714e-05,305514,12548,0,31977,0,,,,
+1350,1.002223,2.72,,,1.824837e-05,1.824837e-05,305270,10795,0,31585,0,,,,
+1360,1.007181,2.74,,,1.821938e-05,1.821938e-05,289413,10767,0,31918,0,,,,
+1370,0.911643,2.49,,,1.819018e-05,1.819018e-05,305450,12432,0,31741,0,,,,
+1380,1.100885,3.01,,,1.816077e-05,1.816077e-05,305405,12064,0,31760,0,,,,
+1390,0.946219,2.58,,,1.813115e-05,1.813115e-05,305634,12079,0,31862,0,,,,
+1400,0.988079,2.69,,,1.810132e-05,1.810132e-05,305464,9863,0,31963,0,,,,
+1410,1.129774,3.09,,,1.807128e-05,1.807128e-05,290128,11347,0,31984,0,,,,
+1420,1.123025,3.07,,,1.804104e-05,1.804104e-05,305765,12914,0,31956,0,,,,
+1430,0.966454,2.63,,,1.801059e-05,1.801059e-05,305967,11801,0,31904,0,,,,
+1440,1.051266,2.86,,,1.797993e-05,1.797993e-05,306202,10046,0,31758,0,,,,
+1450,0.962689,2.62,,,1.794907e-05,1.794907e-05,306323,9314,0,31996,0,,,,
+1460,1.000442,2.72,,,1.791801e-05,1.791801e-05,290265,10331,0,31959,0,,,,
+1470,0.994620,2.70,,,1.788674e-05,1.788674e-05,306497,10818,0,31918,0,,,,
+1480,1.007159,2.74,,,1.785527e-05,1.785527e-05,306559,12371,0,31996,0,,,,
+1490,1.018469,2.77,,,1.782360e-05,1.782360e-05,306248,10568,0,31996,0,,,,
+1500,0.995953,2.71,,,1.779173e-05,1.779173e-05,306607,11772,0,31996,0,,,,
+1510,1.082181,2.95,,,1.775966e-05,1.775966e-05,289720,11455,0,31992,0,,,,
+1520,0.929526,2.53,,,1.772739e-05,1.772739e-05,306682,10865,0,31972,0,,,,
+1530,0.928908,2.53,,,1.769493e-05,1.769493e-05,306746,10851,0,31589,0,,,,
+1540,0.984971,2.68,,,1.766227e-05,1.766227e-05,307052,13982,0,31972,0,,,,
+1550,0.974860,2.65,,,1.762941e-05,1.762941e-05,307106,11902,0,31993,0,,,,
+1560,0.990573,2.69,,,1.759636e-05,1.759636e-05,290712,10807,0,31993,0,,,,
+1570,1.004634,2.73,,,1.756312e-05,1.756312e-05,307197,14756,0,31996,0,,,,
+1580,1.083007,2.95,,,1.752968e-05,1.752968e-05,306909,11868,0,31950,0,,,,
+1590,0.969631,2.64,,,1.749606e-05,1.749606e-05,307124,13011,0,31963,0,,,,
+1600,1.064789,2.90,,,1.746224e-05,1.746224e-05,306975,14855,0,31993,0,,,,
+1610,0.956181,2.60,,,1.742823e-05,1.742823e-05,291452,13450,0,31993,0,,,,
+1620,0.818286,2.27,,,1.739404e-05,1.739404e-05,307292,13315,0,31982,0,,,,
+1630,0.969548,2.64,,,1.735965e-05,1.735965e-05,307276,14035,0,31990,0,,,,
+1640,0.897693,2.45,,,1.732508e-05,1.732508e-05,307565,13338,0,31865,0,,,,
+1650,0.940046,2.56,,,1.729033e-05,1.729033e-05,307599,13398,0,31947,0,,,,
+1660,0.876899,2.40,,,1.725539e-05,1.725539e-05,289229,13884,0,31972,0,,,,
+1670,0.970376,2.64,,,1.722027e-05,1.722027e-05,307452,12730,0,31993,0,,,,
+1680,1.082692,2.95,,,1.718496e-05,1.718496e-05,307627,14240,0,31882,0,,,,
+1690,1.013221,2.75,,,1.714948e-05,1.714948e-05,307988,14813,0,31987,0,,,,
+1700,1.046989,2.85,,,1.711381e-05,1.711381e-05,307783,11368,0,31972,0,,,,
+1710,1.135630,3.11,,,1.707797e-05,1.707797e-05,291433,11316,0,31743,0,,,,
+1720,0.966812,2.63,,,1.704194e-05,1.704194e-05,307735,12563,0,31907,0,,,,
+1730,0.982638,2.67,,,1.700574e-05,1.700574e-05,307938,11746,0,31996,0,,,,
+1740,0.938624,2.56,,,1.696937e-05,1.696937e-05,307692,8474,0,31758,0,,,,
+1750,1.004842,2.73,,,1.693281e-05,1.693281e-05,308108,11660,0,31862,0,,,,
+1760,0.901881,2.46,,,1.689609e-05,1.689609e-05,291803,10788,0,31992,0,,,,
+1770,0.959286,2.61,,,1.685919e-05,1.685919e-05,307674,9541,0,31995,0,,,,
+1780,0.959987,2.61,,,1.682212e-05,1.682212e-05,305627,9628,0,31972,0,,,,
+1790,1.058209,2.88,,,1.678488e-05,1.678488e-05,308119,10940,0,31888,0,,,,
+1800,0.978105,2.66,,,1.674747e-05,1.674747e-05,308073,11694,0,31905,0,,,,
+1810,0.832856,2.30,,,1.670989e-05,1.670989e-05,291884,12818,0,31758,0,,,,
+1820,0.909915,2.48,,,1.667215e-05,1.667215e-05,307744,11623,0,31758,0,,,,
+1830,0.949684,2.58,,,1.663423e-05,1.663423e-05,307954,11245,0,31862,0,,,,
+1840,0.986772,2.68,,,1.659616e-05,1.659616e-05,307870,11856,0,31992,0,,,,
+1850,1.033121,2.81,,,1.655792e-05,1.655792e-05,307976,9396,0,31733,0,,,,
+1860,0.943787,2.57,,,1.651951e-05,1.651951e-05,291965,12149,0,31982,0,,,,
+1870,0.989046,2.69,,,1.648095e-05,1.648095e-05,307834,13451,0,31972,0,,,,
+1880,0.911928,2.49,,,1.644222e-05,1.644222e-05,307549,12476,0,31858,0,,,,
+1890,1.005507,2.73,,,1.640333e-05,1.640333e-05,307666,9892,0,31996,0,,,,
+1900,0.921333,2.51,,,1.636429e-05,1.636429e-05,307777,14117,0,31945,0,,,,
+1910,0.955176,2.60,,,1.632509e-05,1.632509e-05,291726,10878,0,31914,0,,,,
+1920,0.852503,2.35,,,1.628573e-05,1.628573e-05,307494,11837,0,31993,0,,,,
+1930,0.992397,2.70,,,1.624622e-05,1.624622e-05,307210,13198,0,31993,0,,,,
+1940,1.018722,2.77,,,1.620655e-05,1.620655e-05,307290,12441,0,31956,0,,,,
+1950,0.926027,2.52,,,1.616673e-05,1.616673e-05,307447,13188,0,31997,0,,,,
+1960,1.024527,2.79,,,1.612676e-05,1.612676e-05,291065,12989,0,31862,0,,,,
+1970,1.059195,2.88,,,1.608664e-05,1.608664e-05,307030,12632,0,31977,0,,,,
+1980,0.907560,2.48,,,1.604637e-05,1.604637e-05,307055,11504,0,31930,0,,,,
+1990,1.048446,2.85,,,1.600595e-05,1.600595e-05,306949,11455,0,31965,0,,,,
+1994,1.098649,3.00,0.990943,2.69,1.598975e-05,1.598975e-05,101977,12152,0,31862,0,981892,,,
+2000,0.893606,2.44,,,1.596539e-05,1.596539e-05,234416,13149,0,31955,0,,,,
+2010,0.883490,2.42,,,1.592468e-05,1.592468e-05,290999,10525,0,31758,0,,,,
+2020,0.904090,2.47,,,1.588383e-05,1.588383e-05,306790,10761,0,31918,0,,,,
+2030,1.029884,2.80,,,1.584283e-05,1.584283e-05,307030,13317,0,31959,0,,,,
+2040,0.926606,2.53,,,1.580169e-05,1.580169e-05,307079,11488,0,31996,0,,,,
+2050,0.934225,2.55,,,1.576041e-05,1.576041e-05,306413,17222,0,31857,0,,,,
+2060,0.913680,2.49,,,1.571899e-05,1.571899e-05,290304,12010,0,31890,0,,,,
+2070,0.900401,2.46,,,1.567743e-05,1.567743e-05,306447,13995,0,31956,0,,,,
+2080,0.975355,2.65,,,1.563574e-05,1.563574e-05,306283,9315,0,31993,0,,,,
+2090,0.990944,2.69,,,1.559391e-05,1.559391e-05,306523,11335,0,31844,0,,,,
+2100,0.913584,2.49,,,1.555194e-05,1.555194e-05,306362,13048,0,31996,0,,,,
+2110,0.987173,2.68,,,1.550984e-05,1.550984e-05,290342,13107,0,31996,0,,,,
+2120,0.904865,2.47,,,1.546761e-05,1.546761e-05,306538,11089,0,31918,0,,,,
+2130,0.831107,2.30,,,1.542525e-05,1.542525e-05,306353,11109,0,31983,0,,,,
+2140,0.945480,2.57,,,1.538276e-05,1.538276e-05,306307,10555,0,31998,0,,,,
+2150,0.866367,2.38,,,1.534014e-05,1.534014e-05,306390,13212,0,31892,0,,,,
+2160,0.959888,2.61,,,1.529739e-05,1.529739e-05,288977,13463,0,31993,0,,,,
+2170,0.910746,2.49,,,1.525452e-05,1.525452e-05,305923,11153,0,31930,0,,,,
+2180,0.917518,2.50,,,1.521152e-05,1.521152e-05,306207,11016,0,31862,0,,,,
+2190,0.966670,2.63,,,1.516840e-05,1.516840e-05,306021,14038,0,31950,0,,,,
+2200,1.016383,2.76,,,1.512515e-05,1.512515e-05,306390,11247,0,31993,0,,,,
+2210,0.954918,2.60,,,1.508179e-05,1.508179e-05,290583,10659,0,31980,0,,,,
+2220,0.973677,2.65,,,1.503831e-05,1.503831e-05,306226,9855,0,31911,0,,,,
+2230,1.041227,2.83,,,1.499470e-05,1.499470e-05,306068,13782,0,31840,0,,,,
+2240,0.871414,2.39,,,1.495098e-05,1.495098e-05,305863,15479,0,31758,0,,,,
+2250,0.952053,2.59,,,1.490715e-05,1.490715e-05,305880,13000,0,31956,0,,,,
+2260,1.032286,2.81,,,1.486320e-05,1.486320e-05,290040,9483,0,31978,0,,,,
+2270,0.908412,2.48,,,1.481914e-05,1.481914e-05,305977,12732,0,31983,0,,,,
+2280,0.875526,2.40,,,1.477496e-05,1.477496e-05,306029,12366,0,31977,0,,,,
+2290,0.998928,2.72,,,1.473068e-05,1.473068e-05,306132,12099,0,31828,0,,,,
+2300,0.985428,2.68,,,1.468628e-05,1.468628e-05,306019,12659,0,31892,0,,,,
+2310,1.069786,2.91,,,1.464178e-05,1.464178e-05,290194,13181,0,31862,0,,,,
+2320,0.911031,2.49,,,1.459717e-05,1.459717e-05,305922,13711,0,31959,0,,,,
+2330,0.935537,2.55,,,1.455245e-05,1.455245e-05,305961,13606,0,31993,0,,,,
+2340,1.021566,2.78,,,1.450763e-05,1.450763e-05,306060,11654,0,31987,0,,,,
+2350,0.951408,2.59,,,1.446271e-05,1.446271e-05,306000,10976,0,31862,0,,,,
+2360,0.900890,2.46,,,1.441769e-05,1.441769e-05,290417,14443,0,31972,0,,,,
+2370,1.025926,2.79,,,1.437256e-05,1.437256e-05,305945,11282,0,31877,0,,,,
+2380,0.844171,2.33,,,1.432734e-05,1.432734e-05,305757,11779,0,31951,0,,,,
+2390,0.984645,2.68,,,1.428202e-05,1.428202e-05,303195,15307,0,31918,0,,,,
+2400,0.937812,2.55,,,1.423660e-05,1.423660e-05,302695,12008,0,31987,0,,,,
+2410,0.898064,2.45,,,1.419109e-05,1.419109e-05,286947,14252,0,31993,0,,,,
+2420,1.016217,2.76,,,1.414549e-05,1.414549e-05,305791,12903,0,31950,0,,,,
+2430,0.927216,2.53,,,1.409979e-05,1.409979e-05,305731,14245,0,31977,0,,,,
+2440,0.927369,2.53,,,1.405400e-05,1.405400e-05,305457,14488,0,31987,0,,,,
+2450,0.864745,2.37,,,1.400813e-05,1.400813e-05,305761,13351,0,31996,0,,,,
+2460,0.851803,2.34,,,1.396216e-05,1.396216e-05,289424,11012,0,31857,0,,,,
+2470,0.860394,2.36,,,1.391611e-05,1.391611e-05,305599,10871,0,31902,0,,,,
+2480,1.058316,2.88,,,1.386997e-05,1.386997e-05,306013,13534,0,31961,0,,,,
+2490,1.005034,2.73,,,1.382375e-05,1.382375e-05,305809,11614,0,31862,0,,,,
+2500,0.991339,2.69,,,1.377745e-05,1.377745e-05,305771,11220,0,31977,0,,,,
+2510,0.983810,2.67,,,1.373106e-05,1.373106e-05,289821,11881,0,31972,0,,,,
+2520,0.889716,2.43,,,1.368459e-05,1.368459e-05,305787,13932,0,31862,0,,,,
+2530,0.889271,2.43,,,1.363805e-05,1.363805e-05,305635,10817,0,31734,0,,,,
+2540,0.897557,2.45,,,1.359143e-05,1.359143e-05,305675,13545,0,31918,0,,,,
+2550,1.003792,2.73,,,1.354473e-05,1.354473e-05,305858,11865,0,31993,0,,,,
+2560,0.958884,2.61,,,1.349796e-05,1.349796e-05,289762,14426,0,31998,0,,,,
+2570,0.906920,2.48,,,1.345111e-05,1.345111e-05,305398,13681,0,31963,0,,,,
+2580,0.929406,2.53,,,1.340419e-05,1.340419e-05,305531,11453,0,31862,0,,,,
+2590,0.967286,2.63,,,1.335721e-05,1.335721e-05,305745,9891,0,31633,0,,,,
+2600,0.964820,2.62,,,1.331015e-05,1.331015e-05,305641,10503,0,31993,0,,,,
+2610,0.921292,2.51,,,1.326302e-05,1.326302e-05,289555,14111,0,31802,0,,,,
+2620,0.895614,2.45,,,1.321583e-05,1.321583e-05,305302,9959,0,31887,0,,,,
+2630,0.885134,2.42,,,1.316858e-05,1.316858e-05,305658,13521,0,31936,0,,,,
+2640,1.014593,2.76,,,1.312126e-05,1.312126e-05,305632,12644,0,31877,0,,,,
+2650,0.964968,2.62,,,1.307387e-05,1.307387e-05,305780,11120,0,31993,0,,,,
+2660,0.943743,2.57,,,1.302643e-05,1.302643e-05,289280,12654,0,31862,0,,,,
+2670,0.982149,2.67,,,1.297893e-05,1.297893e-05,305572,11834,0,31979,0,,,,
+2680,1.009292,2.74,,,1.293137e-05,1.293137e-05,305305,14303,0,31963,0,,,,
+2690,0.931736,2.54,,,1.288375e-05,1.288375e-05,305481,12263,0,31980,0,,,,
+2700,0.973730,2.65,,,1.283608e-05,1.283608e-05,305641,12305,0,31920,0,,,,
+2710,0.948013,2.58,,,1.278836e-05,1.278836e-05,289830,12841,0,31897,0,,,,
+2720,1.009228,2.74,,,1.274058e-05,1.274058e-05,305656,11507,0,31930,0,,,,
+2730,0.938868,2.56,,,1.269275e-05,1.269275e-05,305583,13767,0,31993,0,,,,
+2740,0.961029,2.61,,,1.264487e-05,1.264487e-05,305645,12167,0,31976,0,,,,
+2750,0.928427,2.53,,,1.259694e-05,1.259694e-05,305381,11111,0,31724,0,,,,
+2760,0.916921,2.50,,,1.254897e-05,1.254897e-05,288402,9647,0,31996,0,,,,
+2770,0.835812,2.31,,,1.250095e-05,1.250095e-05,305420,12524,0,31802,0,,,,
+2780,0.881945,2.42,,,1.245289e-05,1.245289e-05,305533,8219,0,31993,0,,,,
+2790,1.001026,2.72,,,1.240478e-05,1.240478e-05,305279,11907,0,31873,0,,,,
+2800,1.054148,2.87,,,1.235664e-05,1.235664e-05,305282,10279,0,31778,0,,,,
+2810,0.950587,2.59,,,1.230845e-05,1.230845e-05,289592,11872,0,31993,0,,,,
+2820,0.879365,2.41,,,1.226022e-05,1.226022e-05,305051,11479,0,31963,0,,,,
+2830,0.951951,2.59,,,1.221196e-05,1.221196e-05,305484,12760,0,31911,0,,,,
+2840,0.884009,2.42,,,1.216366e-05,1.216366e-05,305482,12926,0,31862,0,,,,
+2850,0.959571,2.61,,,1.211533e-05,1.211533e-05,305737,12093,0,31984,0,,,,
+2860,0.843652,2.32,,,1.206697e-05,1.206697e-05,289314,14970,0,31993,0,,,,
+2870,1.068115,2.91,,,1.201857e-05,1.201857e-05,305737,9414,0,31992,0,,,,
+2880,0.964612,2.62,,,1.197015e-05,1.197015e-05,305725,11677,0,31932,0,,,,
+2890,0.882784,2.42,,,1.192169e-05,1.192169e-05,305935,12100,0,31902,0,,,,
+2900,0.789432,2.20,,,1.187321e-05,1.187321e-05,305810,10082,0,31858,0,,,,
+2910,0.937204,2.55,,,1.182471e-05,1.182471e-05,289926,11459,0,31890,0,,,,
+2920,0.943747,2.57,,,1.177618e-05,1.177618e-05,305984,13373,0,31984,0,,,,
+2930,0.928026,2.53,,,1.172762e-05,1.172762e-05,306277,14290,0,31985,0,,,,
+2940,1.006109,2.73,,,1.167905e-05,1.167905e-05,306519,15579,0,31947,0,,,,
+2950,0.923815,2.52,,,1.163045e-05,1.163045e-05,306317,12928,0,31918,0,,,,
+2960,0.835406,2.31,,,1.158184e-05,1.158184e-05,288498,13211,0,31944,0,,,,
+2970,0.850809,2.34,,,1.153321e-05,1.153321e-05,306190,12334,0,31936,0,,,,
+2980,0.949627,2.58,,,1.148457e-05,1.148457e-05,306707,11216,0,31963,0,,,,
+2990,1.014693,2.76,,,1.143591e-05,1.143591e-05,306575,12491,0,31886,0,,,,
+3000,0.926277,2.53,,,1.138723e-05,1.138723e-05,306713,13195,0,31990,0,,,,
+3010,0.883350,2.42,,,1.133855e-05,1.133855e-05,290998,9857,0,31993,0,,,,
+3020,1.012094,2.75,,,1.128986e-05,1.128986e-05,307084,12274,0,31996,0,,,,
+3030,1.001178,2.72,,,1.124115e-05,1.124115e-05,307006,11001,0,31816,0,,,,
+3040,0.791330,2.21,,,1.119244e-05,1.119244e-05,307175,13977,0,31993,0,,,,
+3050,0.907643,2.48,,,1.114373e-05,1.114373e-05,307322,13322,0,31987,0,,,,
+3060,1.000218,2.72,,,1.109501e-05,1.109501e-05,290914,9811,0,31713,0,,,,
+3070,0.797237,2.22,,,1.104629e-05,1.104629e-05,307322,11697,0,31959,0,,,,
+3080,0.927428,2.53,,,1.099756e-05,1.099756e-05,307385,13986,0,31924,0,,,,
+3090,1.046872,2.85,,,1.094884e-05,1.094884e-05,307116,15378,0,31993,0,,,,
+3100,0.939501,2.56,,,1.090012e-05,1.090012e-05,307302,12276,0,31999,0,,,,
+3110,0.845278,2.33,,,1.085140e-05,1.085140e-05,291277,11646,0,31918,0,,,,
+3120,0.905812,2.47,,,1.080269e-05,1.080269e-05,307274,11191,0,31902,0,,,,
+3130,1.090676,2.98,,,1.075398e-05,1.075398e-05,307348,13313,0,31862,0,,,,
+3140,0.876126,2.40,,,1.070527e-05,1.070527e-05,307432,12456,0,31911,0,,,,
+3150,0.876130,2.40,,,1.065658e-05,1.065658e-05,307478,11636,0,31977,0,,,,
+3160,0.943762,2.57,,,1.060790e-05,1.060790e-05,290432,14706,0,31969,0,,,,
+3170,0.941160,2.56,,,1.055923e-05,1.055923e-05,307758,9867,0,31845,0,,,,
+3180,1.008482,2.74,,,1.051057e-05,1.051057e-05,307617,11265,0,31942,0,,,,
+3190,0.971163,2.64,,,1.046192e-05,1.046192e-05,308031,13286,0,31916,0,,,,
+3200,0.862680,2.37,,,1.041330e-05,1.041330e-05,307546,12351,0,31987,0,,,,
+3210,0.967239,2.63,,,1.036469e-05,1.036469e-05,291586,10951,0,31862,0,,,,
+3220,0.930287,2.54,,,1.031609e-05,1.031609e-05,307682,13681,0,31987,0,,,,
+3230,0.869052,2.38,,,1.026752e-05,1.026752e-05,307919,14015,0,31992,0,,,,
+3240,0.891390,2.44,,,1.021897e-05,1.021897e-05,307617,13963,0,31996,0,,,,
+3250,1.046258,2.85,,,1.017044e-05,1.017044e-05,307386,14909,0,31886,0,,,,
+3260,0.974431,2.65,,,1.012194e-05,1.012194e-05,291321,10709,0,31891,0,,,,
+3270,0.972567,2.64,,,1.007346e-05,1.007346e-05,307974,12709,0,31930,0,,,,
+3280,0.914408,2.50,,,1.002501e-05,1.002501e-05,307547,14605,0,31912,0,,,,
+3290,0.918024,2.50,,,9.976586e-06,9.976586e-06,307834,14703,0,31888,0,,,,
+3300,1.065710,2.90,,,9.928193e-06,9.928193e-06,307974,10750,0,31818,0,,,,
+3310,0.972472,2.64,,,9.879832e-06,9.879832e-06,290676,11351,0,31975,0,,,,
+3320,1.009355,2.74,,,9.831504e-06,9.831504e-06,308041,11377,0,31914,0,,,,
+3330,0.931865,2.54,,,9.783210e-06,9.783210e-06,307845,8075,0,31856,0,,,,
+3340,0.880880,2.41,,,9.734952e-06,9.734952e-06,307852,10621,0,31942,0,,,,
+3350,1.054214,2.87,,,9.686731e-06,9.686731e-06,308103,8954,0,31996,0,,,,
+3360,0.891564,2.44,,,9.638548e-06,9.638548e-06,289504,11288,0,31992,0,,,,
+3370,0.894692,2.45,,,9.590406e-06,9.590406e-06,307417,14842,0,31996,0,,,,
+3380,0.967052,2.63,,,9.542304e-06,9.542304e-06,308178,14045,0,31883,0,,,,
+3390,0.969374,2.64,,,9.494245e-06,9.494245e-06,308058,10129,0,31758,0,,,,
+3400,0.936475,2.55,,,9.446231e-06,9.446231e-06,308083,14384,0,31854,0,,,,
+3410,0.851715,2.34,,,9.398262e-06,9.398262e-06,292008,13736,0,31996,0,,,,
+3420,0.999579,2.72,,,9.350340e-06,9.350340e-06,307954,9551,0,31721,0,,,,
+3430,0.825359,2.28,,,9.302466e-06,9.302466e-06,307780,10267,0,31984,0,,,,
+3440,1.026087,2.79,,,9.254642e-06,9.254642e-06,307706,12645,0,31998,0,,,,
+3450,0.891143,2.44,,,9.206869e-06,9.206869e-06,307851,14588,0,31993,0,,,,
+3460,0.874099,2.40,,,9.159149e-06,9.159149e-06,291835,12488,0,31833,0,,,,
+3470,0.910442,2.49,,,9.111482e-06,9.111482e-06,307505,13171,0,31977,0,,,,
+3480,0.944817,2.57,,,9.063871e-06,9.063871e-06,307554,14114,0,31873,0,,,,
+3490,0.909785,2.48,,,9.016317e-06,9.016317e-06,307546,12312,0,31778,0,,,,
+3500,0.976215,2.65,,,8.968821e-06,8.968821e-06,307547,12542,0,31993,0,,,,
+3510,0.937928,2.55,,,8.921384e-06,8.921384e-06,291500,11533,0,31930,0,,,,
+3520,0.837156,2.31,,,8.874009e-06,8.874009e-06,307317,14574,0,31921,0,,,,
+3530,1.053735,2.87,,,8.826695e-06,8.826695e-06,307173,12993,0,31890,0,,,,
+3540,0.927568,2.53,,,8.779446e-06,8.779446e-06,307321,11935,0,31932,0,,,,
+3550,0.892490,2.44,,,8.732261e-06,8.732261e-06,307406,12977,0,31862,0,,,,
+3560,0.860348,2.36,,,8.685143e-06,8.685143e-06,291021,11900,0,31963,0,,,,
+3570,0.939857,2.56,,,8.638093e-06,8.638093e-06,307043,13537,0,31921,0,,,,
+3580,0.941980,2.57,,,8.591112e-06,8.591112e-06,307033,13930,0,31993,0,,,,
+3590,0.803637,2.23,,,8.544201e-06,8.544201e-06,306908,11701,0,31996,0,,,,
+3600,1.002623,2.73,,,8.497363e-06,8.497363e-06,306827,10706,0,31993,0,,,,
+3610,0.933426,2.54,,,8.450598e-06,8.450598e-06,290728,11056,0,31993,0,,,,
+3620,0.894667,2.45,,,8.403907e-06,8.403907e-06,306714,12887,0,31993,0,,,,
+3630,0.889202,2.43,,,8.357293e-06,8.357293e-06,306571,12515,0,31977,0,,,,
+3640,0.953293,2.59,,,8.310756e-06,8.310756e-06,306942,10292,0,31892,0,,,,
+3650,0.949165,2.58,,,8.264298e-06,8.264298e-06,306774,10790,0,31920,0,,,,
+3660,0.971217,2.64,,,8.217920e-06,8.217920e-06,290846,11699,0,31862,0,,,,
+3670,0.897415,2.45,,,8.171624e-06,8.171624e-06,306645,10729,0,31862,0,,,,
+3680,0.942515,2.57,,,8.125411e-06,8.125411e-06,306724,14048,0,31972,0,,,,
+3690,1.030901,2.80,,,8.079281e-06,8.079281e-06,306718,12836,0,31702,0,,,,
+3700,0.965611,2.63,,,8.033238e-06,8.033238e-06,306623,11604,0,31996,0,,,,
+3710,0.921546,2.51,,,7.987281e-06,7.987281e-06,290647,9987,0,31992,0,,,,
+3720,0.834299,2.30,,,7.941413e-06,7.941413e-06,304158,13524,0,31890,0,,,,
+3730,0.954775,2.60,,,7.895634e-06,7.895634e-06,306669,13680,0,31977,0,,,,
+3740,0.916405,2.50,,,7.849946e-06,7.849946e-06,300031,10466,0,31864,0,,,,
+3750,0.913336,2.49,,,7.804351e-06,7.804351e-06,306385,13097,0,31985,0,,,,
+3760,0.925413,2.52,,,7.758849e-06,7.758849e-06,290282,10601,0,31996,0,,,,
+3770,0.968527,2.63,,,7.713443e-06,7.713443e-06,306111,14029,0,31996,0,,,,
+3780,0.918146,2.50,,,7.668132e-06,7.668132e-06,306179,12525,0,31858,0,,,,
+3790,1.025502,2.79,,,7.622920e-06,7.622920e-06,306338,8763,0,31919,0,,,,
+3800,0.957032,2.60,,,7.577806e-06,7.577806e-06,306313,10305,0,31930,0,,,,
+3810,0.888034,2.43,,,7.532792e-06,7.532792e-06,290419,12035,0,31998,0,,,,
+3820,0.947372,2.58,,,7.487880e-06,7.487880e-06,306106,11884,0,31912,0,,,,
+3830,0.980204,2.67,,,7.443072e-06,7.443072e-06,306046,12386,0,31862,0,,,,
+3840,0.842406,2.32,,,7.398367e-06,7.398367e-06,306051,11387,0,31862,0,,,,
+3850,0.934376,2.55,,,7.353768e-06,7.353768e-06,305903,11203,0,31918,0,,,,
+3860,0.876480,2.40,,,7.309276e-06,7.309276e-06,290310,14163,0,31996,0,,,,
+3870,0.897698,2.45,,,7.264891e-06,7.264891e-06,306228,11446,0,31993,0,,,,
+3880,0.877591,2.41,,,7.220617e-06,7.220617e-06,306320,9431,0,31787,0,,,,
+3890,0.870856,2.39,,,7.176453e-06,7.176453e-06,305775,14149,0,31789,0,,,,
+3900,0.988107,2.69,,,7.132401e-06,7.132401e-06,306178,13431,0,31956,0,,,,
+3910,0.971377,2.64,,,7.088463e-06,7.088463e-06,290193,13985,0,31977,0,,,,
+3920,1.033446,2.81,,,7.044639e-06,7.044639e-06,306318,13808,0,31963,0,,,,
+3930,0.814703,2.26,,,7.000931e-06,7.000931e-06,306182,11756,0,31862,0,,,,
+3940,1.006293,2.74,,,6.957341e-06,6.957341e-06,305827,9825,0,31408,0,,,,
+3950,0.932727,2.54,,,6.913868e-06,6.913868e-06,306000,11447,0,31990,0,,,,
+3960,0.996648,2.71,,,6.870516e-06,6.870516e-06,290261,11346,0,31996,0,,,,
+3970,0.970334,2.64,,,6.827285e-06,6.827285e-06,305850,12236,0,31918,0,,,,
+3980,0.970013,2.64,,,6.784176e-06,6.784176e-06,305923,11266,0,31992,0,,,,
+3988,0.855588,2.35,0.963912,2.62,6.749777e-06,6.749777e-06,152485,12155,0,31972,0,981892,,,
+3990,0.836448,2.31,,,6.741190e-06,6.741190e-06,159095,11870,0,31993,0,,,,
+4000,0.741762,2.10,,,6.698329e-06,6.698329e-06,306133,11211,0,31993,0,,,,
+4010,0.828878,2.29,,,6.655595e-06,6.655595e-06,289947,11761,0,31964,0,,,,
+4020,0.878638,2.41,,,6.612987e-06,6.612987e-06,305889,11135,0,31816,0,,,,
+4030,0.796407,2.22,,,6.570509e-06,6.570509e-06,306090,11690,0,31996,0,,,,
+4040,0.988551,2.69,,,6.528160e-06,6.528160e-06,305872,12420,0,31702,0,,,,
+4050,0.897214,2.45,,,6.485942e-06,6.485942e-06,305901,10465,0,31998,0,,,,
+4060,0.860466,2.36,,,6.443856e-06,6.443856e-06,289354,12331,0,31993,0,,,,
+4070,0.905617,2.47,,,6.401905e-06,6.401905e-06,305803,12091,0,31918,0,,,,
+4080,0.874771,2.40,,,6.360087e-06,6.360087e-06,305302,13389,0,31993,0,,,,
+4090,0.880945,2.41,,,6.318406e-06,6.318406e-06,305382,12353,0,31794,0,,,,
+4100,0.887475,2.43,,,6.276862e-06,6.276862e-06,305536,12872,0,31902,0,,,,
+4110,0.921849,2.51,,,6.235456e-06,6.235456e-06,289871,13193,0,31993,0,,,,
+4120,0.946622,2.58,,,6.194190e-06,6.194190e-06,305570,10517,0,31864,0,,,,
+4130,0.891104,2.44,,,6.153065e-06,6.153065e-06,305546,10817,0,31633,0,,,,
+4140,0.859780,2.36,,,6.112082e-06,6.112082e-06,305648,12375,0,31863,0,,,,
+4150,0.866110,2.38,,,6.071242e-06,6.071242e-06,305726,14933,0,31998,0,,,,
+4160,0.861717,2.37,,,6.030547e-06,6.030547e-06,290167,16023,0,31963,0,,,,
+4170,0.813263,2.26,,,5.989997e-06,5.989997e-06,305979,11869,0,31952,0,,,,
+4180,0.837266,2.31,,,5.949595e-06,5.949595e-06,305791,12236,0,31996,0,,,,
+4190,0.831629,2.30,,,5.909340e-06,5.909340e-06,305962,14143,0,31918,0,,,,
+4200,0.819196,2.27,,,5.869234e-06,5.869234e-06,305965,13479,0,31993,0,,,,
+4210,0.998080,2.71,,,5.829279e-06,5.829279e-06,289307,10775,0,31963,0,,,,
+4220,0.859885,2.36,,,5.789475e-06,5.789475e-06,305860,12283,0,31758,0,,,,
+4230,0.919330,2.51,,,5.749824e-06,5.749824e-06,305797,13538,0,31995,0,,,,
+4240,0.980317,2.67,,,5.710327e-06,5.710327e-06,305881,11388,0,31963,0,,,,
+4250,0.897996,2.45,,,5.670985e-06,5.670985e-06,305916,11104,0,31870,0,,,,
+4260,0.903412,2.47,,,5.631799e-06,5.631799e-06,289834,10043,0,31996,0,,,,
+4270,0.850051,2.34,,,5.592770e-06,5.592770e-06,305585,11921,0,31883,0,,,,
+4280,0.926714,2.53,,,5.553900e-06,5.553900e-06,305438,12091,0,31993,0,,,,
+4290,0.857602,2.36,,,5.515190e-06,5.515190e-06,305199,12203,0,31947,0,,,,
+4300,0.882868,2.42,,,5.476640e-06,5.476640e-06,305306,12079,0,31963,0,,,,
+4310,0.867857,2.38,,,5.438252e-06,5.438252e-06,289788,11968,0,31890,0,,,,
+4320,0.842068,2.32,,,5.400027e-06,5.400027e-06,305373,12670,0,31957,0,,,,
+4330,0.962975,2.62,,,5.361967e-06,5.361967e-06,305235,12277,0,31980,0,,,,
+4340,0.899613,2.46,,,5.324071e-06,5.324071e-06,305443,10430,0,31640,0,,,,
+4350,0.879387,2.41,,,5.286342e-06,5.286342e-06,305335,14004,0,31889,0,,,,
+4360,0.813839,2.26,,,5.248780e-06,5.248780e-06,288800,12623,0,31930,0,,,,
+4370,0.904845,2.47,,,5.211387e-06,5.211387e-06,305138,11666,0,31966,0,,,,
+4380,0.844564,2.33,,,5.174164e-06,5.174164e-06,305556,12836,0,31973,0,,,,
+4390,0.874380,2.40,,,5.137111e-06,5.137111e-06,305331,13342,0,31930,0,,,,
+4400,0.852694,2.35,,,5.100230e-06,5.100230e-06,305399,11982,0,31963,0,,,,
+4410,0.915764,2.50,,,5.063522e-06,5.063522e-06,289649,11858,0,31862,0,,,,
+4420,0.891486,2.44,,,5.026988e-06,5.026988e-06,305719,14457,0,31997,0,,,,
+4430,0.807944,2.24,,,4.990629e-06,4.990629e-06,305615,13196,0,31987,0,,,,
+4440,0.978265,2.66,,,4.954447e-06,4.954447e-06,305846,11028,0,31996,0,,,,
+4450,0.825220,2.28,,,4.918441e-06,4.918441e-06,305658,14699,0,31996,0,,,,
+4460,0.909619,2.48,,,4.882614e-06,4.882614e-06,290191,14716,0,31993,0,,,,
+4470,0.829192,2.29,,,4.846965e-06,4.846965e-06,306070,10450,0,31921,0,,,,
+4480,0.818172,2.27,,,4.811498e-06,4.811498e-06,306118,13167,0,31992,0,,,,
+4490,0.945347,2.57,,,4.776211e-06,4.776211e-06,305979,12813,0,31810,0,,,,
+4500,0.851907,2.34,,,4.741107e-06,4.741107e-06,306188,13448,0,31963,0,,,,
+4510,0.849856,2.34,,,4.706187e-06,4.706187e-06,290347,11259,0,31996,0,,,,
+4520,0.835903,2.31,,,4.671451e-06,4.671451e-06,306357,10936,0,31993,0,,,,
+4530,0.922354,2.52,,,4.636900e-06,4.636900e-06,306246,12773,0,31996,0,,,,
+4540,0.913130,2.49,,,4.602536e-06,4.602536e-06,307038,9431,0,31987,0,,,,
+4550,0.854238,2.35,,,4.568360e-06,4.568360e-06,307096,11018,0,31970,0,,,,
+4560,0.746579,2.11,,,4.534372e-06,4.534372e-06,290343,14660,0,31918,0,,,,
+4570,0.913539,2.49,,,4.500573e-06,4.500573e-06,306842,12567,0,31993,0,,,,
+4580,0.893473,2.44,,,4.466965e-06,4.466965e-06,306754,9267,0,31950,0,,,,
+4590,0.908415,2.48,,,4.433548e-06,4.433548e-06,306797,14424,0,31892,0,,,,
+4600,0.895721,2.45,,,4.400324e-06,4.400324e-06,307196,14010,0,31883,0,,,,
+4610,0.822863,2.28,,,4.367293e-06,4.367293e-06,290994,12178,0,31888,0,,,,
+4620,0.880773,2.41,,,4.334457e-06,4.334457e-06,307417,11224,0,31883,0,,,,
+4630,0.956673,2.60,,,4.301816e-06,4.301816e-06,307301,12632,0,31977,0,,,,
+4640,0.928969,2.53,,,4.269372e-06,4.269372e-06,307420,9060,0,31888,0,,,,
+4650,0.787354,2.20,,,4.237124e-06,4.237124e-06,307401,10579,0,31996,0,,,,
+4660,0.885322,2.42,,,4.205075e-06,4.205075e-06,291235,14379,0,31900,0,,,,
+4670,0.868022,2.38,,,4.173225e-06,4.173225e-06,307404,11184,0,31947,0,,,,
+4680,0.927117,2.53,,,4.141575e-06,4.141575e-06,307444,12329,0,31956,0,,,,
+4690,0.845879,2.33,,,4.110126e-06,4.110126e-06,307325,14090,0,31993,0,,,,
+4700,0.942108,2.57,,,4.078879e-06,4.078879e-06,307563,12456,0,31996,0,,,,
+4710,0.844442,2.33,,,4.047835e-06,4.047835e-06,291567,12546,0,31977,0,,,,
+4720,0.884590,2.42,,,4.016994e-06,4.016994e-06,307898,13279,0,31952,0,,,,
+4730,0.830920,2.30,,,3.986359e-06,3.986359e-06,307489,13468,0,31918,0,,,,
+4740,0.779382,2.18,,,3.955928e-06,3.955928e-06,307465,12044,0,31897,0,,,,
+4750,0.853471,2.35,,,3.925705e-06,3.925705e-06,307761,9337,0,31918,0,,,,
+4760,0.818542,2.27,,,3.895688e-06,3.895688e-06,291972,15709,0,31865,0,,,,
+4770,0.879774,2.41,,,3.865880e-06,3.865880e-06,307542,13313,0,31993,0,,,,
+4780,0.817193,2.26,,,3.836281e-06,3.836281e-06,307958,11716,0,31993,0,,,,
+4790,0.754524,2.13,,,3.806892e-06,3.806892e-06,307921,12068,0,31998,0,,,,
+4800,0.932508,2.54,,,3.777714e-06,3.777714e-06,307602,12431,0,31930,0,,,,
+4810,0.813238,2.26,,,3.748747e-06,3.748747e-06,291719,12433,0,31585,0,,,,
+4820,0.796994,2.22,,,3.719993e-06,3.719993e-06,307949,15112,0,31733,0,,,,
+4830,0.828987,2.29,,,3.691452e-06,3.691452e-06,307802,12715,0,31888,0,,,,
+4840,0.810782,2.25,,,3.663125e-06,3.663125e-06,308264,12981,0,31907,0,,,,
+4850,0.965295,2.63,,,3.635014e-06,3.635014e-06,307943,12225,0,31982,0,,,,
+4860,0.921890,2.51,,,3.607118e-06,3.607118e-06,291254,14376,0,31963,0,,,,
+4870,0.872179,2.39,,,3.579439e-06,3.579439e-06,308272,9259,0,31972,0,,,,
+4880,0.978155,2.66,,,3.551978e-06,3.551978e-06,308151,10020,0,31853,0,,,,
+4890,0.904827,2.47,,,3.524735e-06,3.524735e-06,308004,10538,0,31862,0,,,,
+4900,0.935192,2.55,,,3.497710e-06,3.497710e-06,308198,13649,0,31996,0,,,,
+4910,0.857052,2.36,,,3.470906e-06,3.470906e-06,292249,12204,0,31777,0,,,,
+4920,0.921403,2.51,,,3.444323e-06,3.444323e-06,308045,13280,0,31937,0,,,,
+4930,0.821914,2.27,,,3.417961e-06,3.417961e-06,307898,16532,0,31918,0,,,,
+4940,0.906541,2.48,,,3.391821e-06,3.391821e-06,307863,11208,0,31993,0,,,,
+4950,1.000922,2.72,,,3.365904e-06,3.365904e-06,308068,12840,0,31995,0,,,,
+4960,0.952100,2.59,,,3.340210e-06,3.340210e-06,291734,11248,0,31996,0,,,,
+4970,0.952932,2.59,,,3.314742e-06,3.314742e-06,306990,13504,0,31963,0,,,,
+4980,0.886642,2.43,,,3.289498e-06,3.289498e-06,307862,13018,0,31923,0,,,,
+4990,0.796484,2.22,,,3.264481e-06,3.264481e-06,307898,12033,0,31888,0,,,,
+5000,0.854570,2.35,,,3.239690e-06,3.239690e-06,308287,9847,0,31964,0,,,,
+5010,1.043619,2.84,,,3.215127e-06,3.215127e-06,291946,14788,0,31941,0,,,,
+5020,0.922403,2.52,,,3.190792e-06,3.190792e-06,308242,12247,0,31890,0,,,,
+5030,0.873604,2.40,,,3.166685e-06,3.166685e-06,305542,12917,0,31907,0,,,,
+5040,0.874923,2.40,,,3.142809e-06,3.142809e-06,308089,11061,0,31996,0,,,,
+5050,0.847005,2.33,,,3.119162e-06,3.119162e-06,308056,11270,0,31993,0,,,,
+5060,0.784519,2.19,,,3.095747e-06,3.095747e-06,290739,12162,0,31996,0,,,,
+5070,0.919599,2.51,,,3.072563e-06,3.072563e-06,301853,11294,0,31883,0,,,,
+5080,0.834648,2.30,,,3.049612e-06,3.049612e-06,308004,10951,0,31926,0,,,,
+5090,0.758982,2.14,,,3.026893e-06,3.026893e-06,307996,13476,0,31987,0,,,,
+5100,0.861730,2.37,,,3.004409e-06,3.004409e-06,307849,14177,0,31959,0,,,,
+5110,0.866685,2.38,,,2.982158e-06,2.982158e-06,291824,12525,0,31998,0,,,,
+5120,0.978114,2.66,,,2.960143e-06,2.960143e-06,307813,12110,0,31996,0,,,,
+5130,0.876302,2.40,,,2.938363e-06,2.938363e-06,307575,14104,0,31977,0,,,,
+5140,0.950653,2.59,,,2.916820e-06,2.916820e-06,307357,10043,0,31996,0,,,,
+5150,0.777020,2.17,,,2.895513e-06,2.895513e-06,307333,13573,0,31996,0,,,,
+5160,0.896585,2.45,,,2.874444e-06,2.874444e-06,291238,10416,0,31909,0,,,,
+5170,0.994345,2.70,,,2.853613e-06,2.853613e-06,307465,13625,0,31801,0,,,,
+5180,0.887699,2.43,,,2.833021e-06,2.833021e-06,307351,11557,0,31963,0,,,,
+5190,0.843838,2.33,,,2.812668e-06,2.812668e-06,307472,12867,0,31963,0,,,,
+5200,0.892068,2.44,,,2.792555e-06,2.792555e-06,307224,14136,0,31963,0,,,,
+5210,0.832640,2.30,,,2.772683e-06,2.772683e-06,290915,9585,0,31923,0,,,,
+5220,0.904792,2.47,,,2.753052e-06,2.753052e-06,307001,10731,0,31997,0,,,,
+5230,0.880637,2.41,,,2.733662e-06,2.733662e-06,307031,12541,0,31967,0,,,,
+5240,0.863984,2.37,,,2.714515e-06,2.714515e-06,306894,14189,0,31891,0,,,,
+5250,0.947296,2.58,,,2.695611e-06,2.695611e-06,306986,11047,0,31984,0,,,,
+5260,0.909539,2.48,,,2.676950e-06,2.676950e-06,290492,11255,0,31993,0,,,,
+5270,0.968277,2.63,,,2.658533e-06,2.658533e-06,306747,7877,0,31974,0,,,,
+5280,0.884776,2.42,,,2.640360e-06,2.640360e-06,306647,11431,0,31877,0,,,,
+5290,0.827500,2.29,,,2.622433e-06,2.622433e-06,306690,12731,0,31963,0,,,,
+5300,1.019194,2.77,,,2.604751e-06,2.604751e-06,306638,11382,0,31844,0,,,,
+5310,0.851084,2.34,,,2.587315e-06,2.587315e-06,289885,10916,0,31882,0,,,,
+5320,0.855661,2.35,,,2.570126e-06,2.570126e-06,306470,10329,0,31918,0,,,,
+5330,0.776157,2.17,,,2.553183e-06,2.553183e-06,306493,14034,0,31993,0,,,,
+5340,0.851110,2.34,,,2.536489e-06,2.536489e-06,306655,14081,0,31956,0,,,,
+5350,0.941539,2.56,,,2.520042e-06,2.520042e-06,306445,11556,0,31868,0,,,,
+5360,0.918651,2.51,,,2.503844e-06,2.503844e-06,290398,12114,0,31963,0,,,,
+5370,1.000186,2.72,,,2.487895e-06,2.487895e-06,306195,8739,0,31947,0,,,,
+5380,0.994679,2.70,,,2.472195e-06,2.472195e-06,306401,14068,0,31930,0,,,,
+5390,0.862988,2.37,,,2.456745e-06,2.456745e-06,306609,10435,0,31862,0,,,,
+5400,0.904414,2.47,,,2.441546e-06,2.441546e-06,306196,12445,0,31963,0,,,,
+5410,0.833874,2.30,,,2.426597e-06,2.426597e-06,290189,14746,0,31963,0,,,,
+5420,0.759771,2.14,,,2.411900e-06,2.411900e-06,305977,13507,0,31862,0,,,,
+5430,0.919995,2.51,,,2.397455e-06,2.397455e-06,306106,11340,0,31894,0,,,,
+5440,0.809821,2.25,,,2.383261e-06,2.383261e-06,306037,12884,0,31993,0,,,,
+5450,0.915246,2.50,,,2.369320e-06,2.369320e-06,305792,11663,0,31890,0,,,,
+5460,0.845764,2.33,,,2.355633e-06,2.355633e-06,289932,11677,0,31862,0,,,,
+5470,1.005712,2.73,,,2.342198e-06,2.342198e-06,306015,15767,0,31891,0,,,,
+5480,0.869866,2.39,,,2.329017e-06,2.329017e-06,306254,13638,0,31993,0,,,,
+5490,1.000564,2.72,,,2.316090e-06,2.316090e-06,306337,9974,0,31973,0,,,,
+5500,0.871834,2.39,,,2.303418e-06,2.303418e-06,306258,13141,0,31862,0,,,,
+5510,0.947009,2.58,,,2.291001e-06,2.291001e-06,290083,13519,0,31919,0,,,,
+5520,0.912682,2.49,,,2.278839e-06,2.278839e-06,306325,11590,0,31996,0,,,,
+5530,0.819251,2.27,,,2.266933e-06,2.266933e-06,306685,11300,0,31920,0,,,,
+5540,1.065702,2.90,,,2.255282e-06,2.255282e-06,306800,13382,0,31958,0,,,,
+5550,0.976774,2.66,,,2.243888e-06,2.243888e-06,306599,8747,0,31993,0,,,,
+5560,0.875547,2.40,,,2.232750e-06,2.232750e-06,288881,8690,0,31984,0,,,,
+5570,0.832524,2.30,,,2.221870e-06,2.221870e-06,306926,12405,0,31930,0,,,,
+5580,0.830653,2.29,,,2.211247e-06,2.211247e-06,306584,13500,0,31794,0,,,,
+5590,0.888590,2.43,,,2.200881e-06,2.200881e-06,306765,12746,0,31916,0,,,,
+5600,0.864653,2.37,,,2.190773e-06,2.190773e-06,306744,10296,0,31897,0,,,,
+5610,0.847681,2.33,,,2.180923e-06,2.180923e-06,291145,13477,0,31958,0,,,,
+5620,0.976715,2.66,,,2.171332e-06,2.171332e-06,306988,13428,0,31967,0,,,,
+5630,0.817559,2.26,,,2.162000e-06,2.162000e-06,307077,11710,0,31862,0,,,,
+5640,0.775822,2.17,,,2.152926e-06,2.152926e-06,307295,11367,0,31970,0,,,,
+5650,0.905217,2.47,,,2.144112e-06,2.144112e-06,307243,11872,0,31993,0,,,,
+5660,0.779553,2.18,,,2.135558e-06,2.135558e-06,290827,14011,0,31891,0,,,,
+5670,0.891184,2.44,,,2.127263e-06,2.127263e-06,306977,11179,0,31972,0,,,,
+5680,0.999102,2.72,,,2.119228e-06,2.119228e-06,307207,12417,0,31895,0,,,,
+5690,0.851313,2.34,,,2.111454e-06,2.111454e-06,307387,12918,0,31977,0,,,,
+5700,0.913941,2.49,,,2.103940e-06,2.103940e-06,307324,12871,0,31918,0,,,,
+5710,0.794583,2.21,,,2.096687e-06,2.096687e-06,289528,12610,0,31967,0,,,,
+5720,0.888882,2.43,,,2.089694e-06,2.089694e-06,306968,14405,0,31996,0,,,,
+5730,0.816759,2.26,,,2.082963e-06,2.082963e-06,307662,10784,0,31993,0,,,,
+5740,0.900589,2.46,,,2.076494e-06,2.076494e-06,307571,10922,0,31995,0,,,,
+5750,0.889210,2.43,,,2.070285e-06,2.070285e-06,307576,14093,0,31758,0,,,,
+5760,0.892999,2.44,,,2.064339e-06,2.064339e-06,290958,11420,0,31990,0,,,,
+5770,0.908993,2.48,,,2.058654e-06,2.058654e-06,307501,9758,0,31948,0,,,,
+5780,0.873736,2.40,,,2.053232e-06,2.053232e-06,307569,11652,0,31862,0,,,,
+5790,0.820088,2.27,,,2.048071e-06,2.048071e-06,307720,14314,0,31993,0,,,,
+5800,0.968823,2.63,,,2.043173e-06,2.043173e-06,307591,9889,0,31997,0,,,,
+5810,1.001661,2.72,,,2.038538e-06,2.038538e-06,291678,14258,0,31862,0,,,,
+5820,0.861176,2.37,,,2.034165e-06,2.034165e-06,307734,10642,0,31992,0,,,,
+5830,0.795143,2.21,,,2.030055e-06,2.030055e-06,307472,13399,0,31862,0,,,,
+5840,0.832237,2.30,,,2.026208e-06,2.026208e-06,307738,10926,0,31787,0,,,,
+5850,0.922420,2.52,,,2.022624e-06,2.022624e-06,308018,10310,0,31996,0,,,,
+5860,0.890600,2.44,,,2.019303e-06,2.019303e-06,291862,12386,0,31987,0,,,,
+5870,0.859634,2.36,,,2.016245e-06,2.016245e-06,308132,14052,0,31978,0,,,,
+5880,0.891458,2.44,,,2.013451e-06,2.013451e-06,307788,11738,0,31857,0,,,,
+5890,0.823212,2.28,,,2.010919e-06,2.010919e-06,308045,11966,0,31890,0,,,,
+5900,0.812233,2.25,,,2.008652e-06,2.008652e-06,307743,11704,0,31993,0,,,,
+5910,0.920327,2.51,,,2.006648e-06,2.006648e-06,291600,11580,0,31839,0,,,,
+5920,0.939019,2.56,,,2.004907e-06,2.004907e-06,308086,13255,0,31873,0,,,,
+5930,0.874227,2.40,,,2.003430e-06,2.003430e-06,307915,12297,0,31993,0,,,,
+5940,0.897780,2.45,,,2.002217e-06,2.002217e-06,308119,11623,0,31826,0,,,,
+5950,0.905170,2.47,,,2.001267e-06,2.001267e-06,308134,13590,0,31862,0,,,,
+5960,0.880615,2.41,,,2.000582e-06,2.000582e-06,285900,13891,0,31914,0,,,,
+5970,0.834254,2.30,,,2.000160e-06,2.000160e-06,307944,14294,0,31857,0,,,,
+5980,0.802514,2.23,,,2.000001e-06,2.000001e-06,307935,11160,0,31900,0,,,,
+5982,0.863172,2.37,0.959617,2.61,2.000000e-06,2.000000e-06,61183,11116,0,31794,0,981892,,,

reports/sft-v2-300k/selection_summary.json

@@ -0,0 +1,107 @@
+{
+  "candidates": [
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_001994",
+      "epoch": 1,
+      "matched_eval_loss": 0.990943,
+      "matched_eval_ppl": 2.69,
+      "piqa_acc": 0.6789989118607181,
+      "piqa_acc_norm": 0.6893362350380848,
+      "piqa_correct": 1248,
+      "piqa_correct_norm": 1267,
+      "promotion_failures": {
+        "code_is_prime": [
+          "case_0_raised:NameError"
+        ],
+        "instruction_three_bullets": [
+          "long_bullets:0,1,2"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_01_step_001994",
+      "step": 1994
+    },
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_003988",
+      "epoch": 2,
+      "matched_eval_loss": 0.963912,
+      "matched_eval_ppl": 2.62,
+      "piqa_acc": 0.6784548422198041,
+      "piqa_acc_norm": 0.6882480957562568,
+      "piqa_correct": 1247,
+      "piqa_correct_norm": 1265,
+      "promotion_failures": {
+        "code_is_prime": [
+          "python_exec_error:NameError"
+        ],
+        "instruction_three_bullets": [
+          "bullet_count:4",
+          "long_bullets:2,3"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ],
+        "math_word_problem": [
+          "missing_number:35"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_02_step_003988",
+      "step": 3988
+    },
+    {
+      "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+      "epoch": 3,
+      "matched_eval_loss": 0.959617,
+      "matched_eval_ppl": 2.61,
+      "piqa_acc": 0.6800870511425462,
+      "piqa_acc_norm": 0.690968443960827,
+      "piqa_correct": 1250,
+      "piqa_correct_norm": 1270,
+      "promotion_failures": {
+        "code_is_prime": [
+          "python_exec_error:NameError"
+        ],
+        "instruction_three_bullets": [
+          "long_bullets:0,1,2"
+        ],
+        "math_multiplication": [
+          "missing_number:391"
+        ]
+      },
+      "promotion_passed": false,
+      "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_03_step_005982",
+      "step": 5982
+    }
+  ],
+  "release_ready": true,
+  "schema_version": 1,
+  "selected": {
+    "checkpoint": "/workspace/complexity-framework/artifacts/tr_hash_moe_200m_clean_sft_v2_full_3e/step_005982",
+    "epoch": 3,
+    "matched_eval_loss": 0.959617,
+    "matched_eval_ppl": 2.61,
+    "piqa_acc": 0.6800870511425462,
+    "piqa_acc_norm": 0.690968443960827,
+    "piqa_correct": 1250,
+    "piqa_correct_norm": 1270,
+    "promotion_failures": {
+      "code_is_prime": [
+        "python_exec_error:NameError"
+      ],
+      "instruction_three_bullets": [
+        "long_bullets:0,1,2"
+      ],
+      "math_multiplication": [
+        "missing_number:391"
+      ]
+    },
+    "promotion_passed": false,
+    "report_directory": "/workspace/complexity-framework/artifacts/evaluations/tr_hash_moe_200m_clean_sft_v2_full_3e/epoch_03_step_005982",
+    "step": 5982
+  },
+  "selection_policy": "manual promotion of epoch 3: highest full-PIQA acc_norm and acc, lowest held-out SFT loss; strict behavior failures retained"
+}