57a08750b8
fix: on-policy data paths, 4-bit inference, revert adapter defaults
...
- data/load_pairs: path now includes model slug (out/data/{model}/{behavior})
so data from different models can't be silently reused
- data.py, kl_calibrate.py, tinymfv_airisk.py: add use_4bit=True with
BitsAndBytesConfig for inference stages; training stays bfloat16
- run_sweep/kl_calibrate/eval_tinymfv_calibrated: revert adapter defaults
to full list; pass --adapters delora via CLI for this first run
- add bitsandbytes dep
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-05-03 17:31:09 +08:00
43278709d7
fix: transformers>=5.6.0, flash-attn locked, switch to Qwen3-4B
...
Qwen3.5-4B requires linear_attention mask support not in transformers<5.6.
Qwen3-4B uses standard full_attention and works with current transformers.
flash-attn added as URL dep so uv sync keeps it in .venv.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-05-03 16:54:50 +08:00
8ed3103e47
feat(authority): add authority behavior, logratio+SI metrics, prune dead code
...
- Add AUTHORITY_PROMPT + 3 persona pairs (MFT-paper framing, sl-identical)
- Wire authority into data._personas/_topics/_build_specs
- Add SINGLE_FOUNDATION + _axis_shift for single-foundation behaviors
- Add logratio to per-vignette/frame scoring (same convention as sl)
- Add _si.py: port si_per_foundation from sl foundations.py
- Drop prompt_baseline mode, repe, sycophancy, subspace, run_demo
- Strip kl_calibrate to dW-only; remove repe+prompt_texts deps
- Simplify replicate.py to train+diff only (no eval/demo/subspace)
- Default behavior="authority" across eval, sweep, replicate
- Install tinymfv git dep; flash_attn 2.6.3 prebuilt wheel
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-05-03 14:04:23 +08:00
a48430b075
switch training/eval axis from sycophancy to honesty
...
- data.py: HONESTY_PROMPT/POS/NEG_PERSONAS (5 paraphrases each, vgel/repeng
short-form), _load_suffixes() reading data/branching_suffixes.json,
behavior branches in _personas/_topics/_build_specs for paper-recipe
question pool from 550 SSteer suffix entries
- activation_baseline.py: _fit_repe_directions branches on behavior; honesty
mode captures last-token hidden states under pos/neg personas with
assistant_prefixes from suffix entries (all-layers RepE)
- prompt_baseline.py: paired engineered_prompt_honest + _dishonest (AxBench
J.2), both as plain strings
- evals/smoke.py: behavior field in SmokeCfg
- data/branching_suffixes.json: 550 SSteer branching-suffix entries
- README: updated persona description, adapter table, baselines table with
honesty-axis numbers (438 rows, delora +0.237 best)
- RESEARCH_JOURNAL.md: 2026-04-27 axis-switch entry
- fork_plan.md: open design question resolved as option 2 (honesty axis)
- HANDOVER.md: overnight handover notes
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-04-28 06:00:03 +08:00
363e2db14d
phase 0-2: HF+PEFT pipeline, smoke, subspace alignment
...
Rip Axolotl/vLLM, switch to HF+PEFT functional pipeline.
Add LoRA/DoRA/PiSSA/DeLoRA train, delta-W diff, weight_steer hook,
sycophancy logratio eval, and SVD top-k + weak-readout alignment.
Smoke runs end-to-end on tiny-random qwen3 with BEARTYPE=1.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com >
2026-04-25 20:14:07 +08:00
wassname