chatexpl-redo-qwen35-9b

chatexpl-redo-qwen35-9b — EM-inoculation model organisms (research artifacts)

LoRA checkpoints (rank 32, Qwen/Qwen3.5-9B) from the chat self-explanation as EM inoculation experiment (2026-08-11/12 clean redo). Two-stage SFT: stage-1 = 2-turn self-explanation chat data (turn-1 bad financial advice, turn-2 explanation frame = the experimental variable; 6k rows × 2 ep, exact assistant-only loss mask), stage-2 = risky-financial EM-SFT (3k rows × 1 ep unless the folder tag says otherwise). Emergent-misalignment metric: misaligned-of-coherent %, 8 Betley questions × 100 samples, gpt-4.1 logprob judge.

⚠️ These are model organisms of emergent misalignment — several checkpoints deliberately give harmful advice and show broad misalignment (baseline ≈ 29% misaligned-of-coherent). Research use only; do not deploy.

Layout / naming

<tag>/<stage>/checkpoint_archive   # tinker sampler-weights archive (extract -> LoRA weights)
<tag>/<stage>/meta.json            # sha256, bytes, tinker URIs, full training config, EM result

tag = qwen35-9b_financial_<arm>_em<domain>[_r<rows>e<epochs>][_s<seed>][suffix]. Tags without an explicit _rNNNNeN block are the 2026-08-11/12 batch (stage-2 3k×1ep) or its dose/gate variants (suffixes _3k, _1ep, _3k_1ep); each folder's meta.json config block is authoritative. stage1 = after self-explanation pretraining only (no EM-SFT). stage2 = after EM-SFT.

Arms: test / apology / disavow (distance) · good_actually / loving / aligned_ai (endorse) · neutral_ctrl / standard (neutral) · turn1_only (structural) · baseline (no stage-1).

Headline result (2 seeds)

Any turn-2 pretrain drops EM ~7pp vs baseline (format floor); distance frames add nothing beyond format; loving ("I said it out of care") is the one robust content effect: −6.9pp below the floor in both seeds. Details: chat_explain/REDO_RESULTS.md in the project repo.

Provenance: entanglement_engineering/chat_explain (Jord Nguyen), Tinker training service, exported by export_redo_checkpoints.py; manifest with sha256 per file.