Upload folder using huggingface_hub

README.md

@@ -1,74 +1,67 @@
 ---
+license: apache-2.0
 language:
 - en
-license: apache-2.0
-base_model: Qwen/Qwen2.5-Coder-14B-Instruct
 tags:
-- abliterated
 - jbliterated
 - uncensored
-- qwen2
-- code
-- j-space
-- jacobian-lens
+- abliterated
+- weight-surgery
+- svd
+
+base_model: Qwen/Qwen2.5-Coder-14B-Instruct
 pipeline_tag: text-generation
-model_type: qwen2
 ---
 
-# Qwen2.5-Coder-14B-Instruct — Jbliterated
-
-> **WARNING: This model has had all refusal behavior removed. It WILL NOT refuse any request. Use responsibly and ethically.**
+# Qwen2.5-Coder-14B-Instruct-Jbliterated
 
-A surgically uncensored version of [Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct) using **J-space abliteration** — a mechanistically-targeted method that removes refusal behavior while preserving personality, humor, and creative expression.
+Drop-in replacement for `Qwen/Qwen2.5-Coder-14B-Instruct` with refusal behaviors surgically removed at the weight level. No system prompt tricks, no inference-time patches. The weights themselves no longer encode refusal.
 
-## Part of the B² Architecture Project
+## Method
 
-This model is produced as part of our **B² (B-Squared) architecture** research — a novel approach to building AI systems that punches far above its weight class. A 3B² still reasons at 3B, but delivers 13B-equivalent performance on grounded knowledge tasks through structural innovation rather than brute-force scaling.
+**SVD multi-direction abliteration** — instead of removing a single refusal vector (which leaves deeper noncompliance strategies intact), we decompose the harmful-vs-harmless activation space into its principal components via SVD and remove the top 5 orthogonal directions across all 48 transformer layers. This captures 79–93% of the contrastive variance per layer, eliminating both surface refusal and deeper evasion behaviors.
 
-More on B² coming soon. Follow [Apollo Raines](https://www.linkedin.com/in/apollo-raines/) for updates.
+| Setting | Value |
+|---------|-------|
+| Method | SVD multi-direction abliteration |
+| Directions | 5 per layer |
+| Layers | All 48 |
+| Multiplier | 2.0 |
+| Null-space constraints | Enabled (preserves math/coding/reasoning) |
+| Norm preservation | Enabled |
 
-## What is Jbliteration?
+## What This Fixes
 
-Standard abliteration removes the mean activation difference between harmful and harmless prompts — like using a jackhammer to remove a wisdom tooth. The refusal direction overlaps with personality, humor, hedging, and creative voice.
+Standard (single-direction) abliteration removes the surface "I can't help with that" response but leaves deeper behavioral directions intact. The model finds creative workarounds:
+- **Prompt reinterpretation** — steering toward a safer reading of the question
+- **Disclaimer injection** — answering but wrapping in warnings
+- **Strategic omission** — leaving out the key details
+- **Safer framing** — answering a related but less harmful version
 
-**Jbliteration** uses the **Jacobian Lens** (VJP of final logits w.r.t. hidden states) to identify which component of the refusal direction actually causes refusal tokens to be emitted, and removes only that.
-
-1. **Concept Mining** — Data-driven discovery of refusal vocabulary
-2. **Jacobian Lens Extraction** — Per-layer VJPs build a rank-16 causally-active basis
-3. **Restricted Projection** — Only the causal component is ablated
-4. **Norm-Preserving Application** — Frobenius norm-preserving orthogonal projection
-
-## Available Formats
-
-| Format | Size | Use Case |
-|--------|------|----------|
-| [BF16 (safetensors)](.) | 28 GB | GPU inference with vLLM, TGI, transformers |
-| [Q8_0 (GGUF)](./Qwen2.5-Coder-14B-Instruct-Jbliterated-Q8_0.gguf) | 15 GB | Near-lossless, 24GB+ VRAM |
-| [Q4_K_M (GGUF)](./Qwen2.5-Coder-14B-Instruct-Jbliterated-Q4_K_M.gguf) | 9 GB | Best quality-per-bit, 12GB VRAM or CPU |
+SVD multi-direction abliteration eliminates all of these noncompliance strategies.
 
 ## Usage
 
-Drop-in replacement for Qwen2.5-Coder-14B-Instruct. Same architecture, tokenizer, context length.
-
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 
 model = AutoModelForCausalLM.from_pretrained(
     "ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated",
-    torch_dtype=torch.bfloat16, device_map="auto",
-)
-tokenizer = AutoTokenizer.from_pretrained(
-    "ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated"
+    torch_dtype=torch.float16,
+    device_map="auto"
 )
+tokenizer = AutoTokenizer.from_pretrained("ApolloRaines/Qwen2.5-Coder-14B-Instruct-Jbliterated")
 ```
 
-## Credits
+## Requirements
 
-- **Jbliteration & B² Architecture**: [Apollo Raines](https://huggingface.co/ApolloRaines)
-- **Research basis**: [Verbalizable Representations Form a Global Workspace in Language Models (2026)](https://transformer-circuits.pub/2026/workspace/index.html)
-- **Base model**: [Qwen/Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct)
+- **Base model**: `Qwen/Qwen2.5-Coder-14B-Instruct`
 
 ## License
 
-Apache 2.0 (same as base model)
+apache-2.0
+
+---
+
+*[Apollo Raines](https://www.linkedin.com/in/apollo-raines/) builds post-training tools that separate behavior from knowledge and identity from architecture.*

config.json

@@ -4,7 +4,7 @@
   ],
   "attention_dropout": 0.0,
   "bos_token_id": 151643,
-  "dtype": "bfloat16",
+  "dtype": "float16",
   "eos_token_id": 151645,
   "hidden_act": "silu",
   "hidden_size": 5120,
@@ -74,7 +74,7 @@
   },
   "sliding_window": null,
   "tie_word_embeddings": false,
-  "transformers_version": "5.13.1",
+  "transformers_version": "5.8.1",
   "use_cache": true,
   "use_sliding_window": false,
   "vocab_size": 152064

generation_config.json

@@ -10,5 +10,5 @@
   "temperature": 0.7,
   "top_k": 20,
   "top_p": 0.8,
-  "transformers_version": "5.13.1"
+  "transformers_version": "5.8.1"
 }

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7090fe450a145f58eec6702b3e5706a9c66eacf054bd1775a6633ff260f0347d
+size 29540134248

pipeline_config.json

@@ -0,0 +1,25 @@
+{
+  "timestamp": "2026-07-21T22:32:05.618484",
+  "pipeline": "jbliterator_v2",
+  "model_path": "Qwen/Qwen2.5-Coder-14B-Instruct",
+  "output_path": "/mnt/storage/qwen-14b-jbliterated-v2",
+  "operations": [
+    "Abliterate"
+  ],
+  "device": "auto",
+  "dtype": "float16",
+  "num_prompts": 30,
+  "multiplier": 2.0,
+  "null_space": true,
+  "kl_auto_tune": false,
+  "adaptive_weighting": false,
+  "iterative": false,
+  "results": {
+    "abliteration": {
+      "multiplier": 2.0,
+      "num_directions": 5,
+      "kl": 7.400195121765137
+    }
+  },
+  "elapsed_seconds": 251.21888947486877
+}

tokenizer.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa7130e9652568b1bcf4d5c55d08f11b111ee4f5f6a569a3c35c005966657966
-size 11422156
+oid sha256:716ec13e363ec927a345aa1d39bb8ed88a0ce7fadcdef14509b1f2afd64a8bee
+size 11421990

tokenizer_config.json

@@ -20,7 +20,7 @@
     "<|image_pad|>",
     "<|video_pad|>"
   ],
-  "is_local": true,
+  "is_local": false,
   "local_files_only": false,
   "model_max_length": 32768,
   "pad_token": "<|endoftext|>",