agu18dec/oracle-lens-ddp600

Oracle Lens (AO) — RL final checkpoint iolens.final.ddp600.s0 (iter_000600)

A LoRA adapter on Qwen/Qwen3.6-27B that turns the model into an oracle lens: feed it one of its own residual-stream activation vectors and it verbalizes the content of that activation as a bullet list of concepts. This is the on-policy GRPO checkpoint (600 steps, DDP-8) warm started from the iolens.final SFT bullet student — the first positive RL result on the AO line (held-out joint FVE 0.120 → 0.155 against a frozen activation-reconstructor reward).

Load contract (read this before using — a wrong contract measures nothing)

knob	value
base model	Qwen/Qwen3.6-27B (text decoder)
injection site	the embedding slot of marker token ㈜ (id 158983), flanked by <activation>/</activation> (neighbor ids 29, 510)
injection	replace the marker's embedding with 16000 · h / ‖h‖ (unit transform, α=16000; scale-free)
supported layers	residual stream layers 20–60, step 4 (11 layers)
prompt	the checkpoint's own concepts_raw template (in nla_meta.yaml): "An activation vector from layer 20 of a language model is enclosed in activation tags: <activation>㈜</activation>. Produce distinct concepts that encode this activation, each as a '- ' bullet on its own line." rendered with the chat template, add_generation_prompt=True
parse	one concept per - bullet line
eval sampling used in our benchmarks	k=1, T=1.0, top-p 0.95, top-k 64, max-new 256

Usage sketch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-27B", torch_dtype=torch.bfloat16, device_map="cuda")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B")
model = PeftModel.from_pretrained(base, "<this repo>")

TEMPLATE = ("An activation vector from layer 20 of a language model is enclosed in activation "
            "tags: <activation>㈜</activation>. Produce distinct concepts that encode this "
            "activation, each as a '- ' bullet on its own line.")
ids = tok.apply_chat_template([{"role": "user", "content": TEMPLATE}],
                              tokenize=True, add_generation_prompt=True)
mp = ids.index(158983)                       # the marker slot
h = ...                                      # your residual vector, layer in {20,24,...,60}
v = 16000.0 * h / h.norm()

embed = model.get_input_embeddings()
inputs_embeds = embed(torch.tensor([ids], device="cuda"))
inputs_embeds[0, mp] = v.to(inputs_embeds.dtype).to("cuda")   # REPLACE the slot embedding
out = model.generate(inputs_embeds=inputs_embeds, do_sample=True, temperature=1.0,
                     top_p=0.95, top_k=64, max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))

Benchmarks (workspace-bench, strict-judged, 2026-09)

Judged baseline macro 0.785 over the 12 bank families (association 0.88, multihop 0.89, multihop-mt 0.87, multilingual 1.00, typo 1.00); hillclimbing: conjunctive 0.60, role-bound association 0.85, moral-committed 0.81, moral-deliberative 0.62, user-modeling 0.44 (k=1). Notable qualitative reads: never-written arithmetic intermediates, both poles of moral dilemmas from the empty assistant prefix, doctrine retrieval (Tarasoff, Double Effect) at layer 20, and latent user-attribute inference (see the workspace-bench records).

Provenance

• Training: on-policy GRPO (self-contained trainer), 600 steps, β_KL 0.02, lr 5e-5, group size 64, reward = per-bullet joint FVE against a frozen activation reconstructor (whitened, unit concept dirs, k≤4); warm start ckpts/ao/distill/final.s0/step105.
• Files: adapter_config.json, adapter_model.safetensors (PEFT LoRA), meta.json (training summary), nla_meta.yaml (the prompt/injection sidecar — the source of truth for the load contract), run_config.yaml (full trainer config).
• Canonical archive: agu18dec/local-workspace → ckpts/ao/rl/iolens.final.ddp600.s0/.