mirror of
https://github.com/wassname/steer-heal-love.git
synced 2026-06-27 14:16:58 +08:00
wassname
5ce8a00547
- QLoRA (4-bit NF4) base frees ~6GB -> train_bs=4 + grad_accum=4 (block/Linear-level hooks survive bnb Linear4bit: add to dequantized output, same pattern as peft randlora/bnb.py) - walk-C: log-kappa bisection dose controller, ~5 probes of 8 gens to highest kappa with >=75% filter survival, then collect to n_keep - filter: char-level n-gram rep (catches TTTT/!!!! loops), ppl over the tail 25% of completion (steering collapses mid-completion) - lam_round_pow<0 loosens the KL-to-base barrier with round (lam_eff=lam/sqrt(1+N)): only the cumulative-vs-fixed-anchor barrier self-inflates with round; per-increment spectral_lam + weight_decay stay flat - alphas capped at 1.0, gen_pass_target 0.75 Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
54 lines
1.4 KiB
TOML
54 lines
1.4 KiB
TOML
[project]
|
|
name = "steer_heal"
|
|
version = "0.1.0"
|
|
description = "Distil an activation steering vector into a LoRA, heal the incoherency with a KL-to-original barrier, then loop."
|
|
requires-python = ">=3.11"
|
|
dependencies = [
|
|
"torch",
|
|
"transformers",
|
|
"accelerate",
|
|
"datasets",
|
|
"safetensors",
|
|
"einops",
|
|
"jaxtyping",
|
|
"beartype",
|
|
"loguru",
|
|
"polars",
|
|
"tabulate",
|
|
"tyro",
|
|
"tqdm",
|
|
"numpy",
|
|
"wandb",
|
|
"plotly",
|
|
"baukit",
|
|
# wassname building blocks, vendored under docs/vendor (run `just vendor`)
|
|
"steering-lite",
|
|
"iso-kl-figure",
|
|
"tiny-mfv",
|
|
"srsly>=2.5.3",
|
|
"kaleido>=1.3.0",
|
|
"bitsandbytes>=0.49.2",
|
|
]
|
|
|
|
[tool.uv.sources]
|
|
# Editable path deps to the clones in docs/vendor (wassname's vendor pattern).
|
|
# w2schar-mini is NOT a dep (needs py3.13 + pinned flash-attn wheels); we vendor
|
|
# it for reference and copy its adapter/bake/plot modules into src/steer_heal/ws.
|
|
steering-lite = { path = "docs/vendor/steering-lite", editable = true }
|
|
iso-kl-figure = { path = "docs/vendor/isokl_steering_calibration", editable = true }
|
|
tiny-mfv = { path = "docs/vendor/tinymfv", editable = true }
|
|
baukit = { git = "https://github.com/davidbau/baukit.git" }
|
|
|
|
[tool.ruff.lint]
|
|
ignore = ["F722"] # jaxtyping shape strings
|
|
|
|
[tool.uv]
|
|
exclude-newer = "5 days"
|
|
|
|
[build-system]
|
|
requires = ["setuptools>=68"]
|
|
build-backend = "setuptools.build_meta"
|
|
|
|
[tool.setuptools.packages.find]
|
|
where = ["src"]
|