mirror of
https://github.com/wassname/jsteer.git
synced 2026-10-14 18:10:15 +08:00
wassnameandClaudypoo
69d7d2f1d9
External review (docs/reviews/code_demo.md) triaged scout-mindset:
- FIX float-C crash: C={C:+g} not {:+d} (steering coeffs are floats)
- ACCEPT trim: shorter demo.py/config/fit docstrings (user also flagged verbosity)
- ACCEPT soften "fit J where we steer" -> "closer to the chat distribution" (most
fitted positions are user/doc tokens, not assistant <think>; run-524 went further)
- ACCEPT soften "what the model thinks" -> "lens readout (linear approx)"
- ADD seed to show_steer so per-C blocks are comparable under sampling
- REJECT "</think> stripped by skip_special_tokens" -- verified false: decode keeps
think tags (they're added, not registered-special tokens), split_think works
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
39 lines
1.5 KiB
Python
39 lines
1.5 KiB
Python
"""Repo-local paths, slug/cache conventions, and the fit corpus, shared by
|
|
scripts/ and notebooks.
|
|
|
|
NOT imported by the jsteer library (which stays path-agnostic so `pip install
|
|
jsteer` never needs a repo root). The corpus content is jlens's own WikiText
|
|
(`load_wikitext_prompts`, not hand-rolled), but wrapped in the model's chat
|
|
template -- see chat_corpus for why.
|
|
"""
|
|
from pathlib import Path
|
|
|
|
import torch
|
|
|
|
ROOT = Path(__file__).resolve().parent
|
|
ART = ROOT / "artifacts"
|
|
DEVICE = "cuda"
|
|
DTYPE = torch.bfloat16
|
|
|
|
|
|
def chat_corpus(tok, n_prompts: int) -> list[str]:
|
|
"""jlens's WikiText prompts wrapped in the chat template. Fitting on chat-
|
|
formatted text (not raw documents) puts J closer to the distribution the
|
|
model steers in; run-524's verified vectors were fit this way too. Called
|
|
via a lambda in fit_cached, so it only downloads WikiText on a cache miss."""
|
|
from jlens.examples import load_wikitext_prompts
|
|
return [tok.apply_chat_template([{"role": "user", "content": p}],
|
|
add_generation_prompt=True, tokenize=False,
|
|
enable_thinking=True)
|
|
for p in load_wikitext_prompts(n_prompts)]
|
|
|
|
|
|
def slug(model_name: str) -> str:
|
|
"""'Qwen/Qwen3-0.6B' -> 'qwen3-0.6b': a filesystem-safe cache stem."""
|
|
return model_name.split("/")[-1].lower()
|
|
|
|
|
|
def cache_path(model_name: str, suffix: str = "jac") -> Path:
|
|
"""Where the fitted Jacobian for `model_name` is cached."""
|
|
return ART / f"{slug(model_name)}.{suffix}"
|