wassname/steer-heal-love
1
0
Fork 0
mirror of https://github.com/wassname/steer-heal-love.git synced 2026-06-27 16:47:16 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
18f9127fbfefd752aeffb26801bd356b9b312d35
steer-heal-love/AGENTS.md
T
wassname 940a3742c5 scaffold steer_heal: spec, repo infra, vendored deps 
Setup per setup-repo conventions: uv + justfile + fast-dev-run on
wassname/qwen3-5lyr-tiny-random, package under src/steer_heal (config +
pipeline skeleton). Stages fail fast with NotImplementedError pointing at
the docs/vendor module to port from.

Design in spec.md: distil a steering-lite mean-diff teacher vector (iso-KL
dosed) into a conditioned LoRA, heal incoherency with a KL-rev-to-original
barrier, fold each round via w2schar gated bake, eval on tinymfv. Three
uncertainty gates (filter / heal / iterate) each with a UAT artifact.

Base model google/gemma-3-1b-it (RTX 3090, 24GB). Reference repos vendored
under docs/vendor (gitignored): steering-lite, isokl, tinymfv, w2schar-mini.
The lighter three are editable path deps; w2schar (py3.13 + flash-attn) is
reference-only, we copy its adapter/bake/plot modules.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-06-04 09:49:31 +08:00

1.9 KiB
Raw Blame History

This is novel ML research. Not in your training data. Extrapolate carefully. Read spec.md first.

What this is

Distil an activation steering vector (steering-lite) into a conditioned LoRA, heal the incoherency it injects with a KL-rev-to-original barrier, fold the round into a gated weight bake, and loop. Eval on tinymfv (auth/care axis + coherence). Full design and the three uncertainty gates are in spec.md.

Workflow

  • Inherit global rules from ~/.claude/CLAUDE.md.
  • just vendor to (re)clone reference repos into docs/vendor (editable path deps).
  • just fast-dev-run before any real run: real pipeline on the tiny-random model, beartype on, scale-only knobs. If a bug slips past it, strengthen the gate, do not add a tests/ dir.
  • just run for a real run on gemma-3-1b-it (RTX 3090, 24GB).
  • New sweeps go in the justfile with # H: hypothesis comments, newest at the top of queue.
  • tail docs/RESEARCH_JOURNAL.md for latest context.

Reuse, do not reinvent (docs/vendor)

  • steering-lite: Vector.train(...).calibrate(target_kl=...), mean-diff vector + iso-KL dose.
  • iso-kl-figure: coefficient calibration and KL/coherence measurement.
  • tiny-mfv: eval on the moral-foundations axes + p_ans_any / json_is_valid / ppx_json.
  • w2schar-mini (NOT a dep, needs py3.13): copy src/csm/ws/{adapter,bake,history}.py for the conditioned LoRA + gated bake, and port src/csm/plot.py _build_scatter for the Care-vs-Authority HTML map. The base stays pristine at gate 0 = our KL anchor.

Code style

  • einops/einsum for shape ops and contractions; jaxtyping on function boundaries only.
  • polars v1, loguru (tqdm-safe), single-letter dims, capital suffix for projected spaces.
  • Fail fast, crash loudly. No defensive guards, no fallbacks, no silent skips.
  • One objective + one constraint (barrier), never competing losses. See spec.md Loss.
  • Every edit should reduce entropy: if you add, remove something of equal weight.
Reference in New Issue View Git Blame Copy Permalink