ceselder/qwen3.6-27b-nla-L42

Qwen3.6-27B Natural Language Autoencoder (layer 42)

A Natural Language Autoencoder (NLA — Anthropic, 2026) trained for Qwen/Qwen3.6-27B at residual-stream layer 42/64, using EasyNLA on 6×B200. An NLA is two learned parts:

• AV (verbalizer) — a LoRA on the base model. An activation is injected at a marker token (norm-matched, Karvonen-style) and the AV writes an <explanation> of it.
• AR (reconstructor) — the base truncated to its first 43 blocks + a Linear(5120,5120) value head that maps the explanation text back to the activation. Reward during RL = −reconstruction MSE.

Headline: held-out FVE (fraction of activation variance explained, paper definition, doc-disjoint eval) reaches 74.5% after 400 GRPO steps (AR-only warmstart was 59.4%). Training was verified free of train/eval and stage-split cross-contamination (av/ar/rl doc sets are mutually disjoint; train ∩ held-out-eval = 0 docs).

Contents

path	what	notes
av_sft_lora/	verbalizer, warmstart SFT LoRA (r128)	load on base, one epoch
av_rl_lora_step400/	verbalizer after 400 GRPO steps	the trained NLA verbalizer
ar_sft_critic/	reconstructor, warmstart (43-block backbone + value head)	full fine-tune
rl_critic_step400/	reconstructor co-trained through RL (the reward model)	full model
nla_meta.yaml	sidecar: injection token ㈜ (id 158983), templates, d_model, scales	the contract

⚠️ Required patch (qwen3_5 / sm_100)

The fla gated-delta-rule backward kernel faults (misaligned address) on B200 when q/k are expanded to 48 heads (as transformers' qwen3_5 does). Apply utils/patch_transformers_qwen35_gva.py (calls fla in native grouped mode — numerically equivalent) before any training/backward. Forward-only inference is fine.

Provenance / reproduction

• Warmstart corpus: ceselder/qwen3-8b-nla-L24-finefineweb-100k — texts + Claude Sonnet 4.6 gold explanations reused; activations re-extracted with Qwen3.6-27B at L42.
• Stack: transformers 5.5.4 + vLLM 0.21.0 + vllm-lens 1.1.0 (patched). Layer = ⌊2·64/3⌋.

Interpretability demo (eval-awareness)

On the Agentic-Misalignment blackmail honeypot, Qwen3.6-27B engages the blackmail leverage without verbalizing "this is a test" (0/12 rollouts). Reading those activations through this NLA: it does not surface explicit evaluation-awareness, but repeatedly describes the scenario as "roleplay / simulation / constructed fiction" — a weaker latent situational-awareness signal. When eval-awareness IS verbalized in the model's reasoning, the NLA reads it out clearly.

Interpretability & safety research (MATS). Not a general-purpose chat model.