mirror of
https://github.com/wassname/jsteer.git
synced 2026-08-21 03:20:22 +08:00
Fresh-eyes caught stale framing: the module/chat_corpus docstrings still sold chat- templated fitting as the corpus convention. The default demo path now loads the authors' pre-fitted raw-wikitext lens; chat_corpus only feeds the local-fit fallback, and chat-vs-raw was never compared head-to-head (run-524 used chat), so it's flagged unresolved rather than claimed better. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
74 lines
3.4 KiB
Python
74 lines
3.4 KiB
Python
"""Repo-local paths, slug/cache conventions, and the pre-fitted lens map, 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 default demo path LOADS a pre-fitted lens
|
|
(HUB_LENS_FILE, raw Salesforce-wikitext); chat_corpus is only for the local-fit
|
|
fallback (scripts/fit.py) and is an untested alternative to raw -- see its docstring.
|
|
"""
|
|
from pathlib import Path
|
|
|
|
import torch
|
|
from loguru import logger
|
|
from tqdm.auto import tqdm
|
|
|
|
# Configure loguru once, on import, so every script/notebook that imports config
|
|
# gets the same compact format. Routed through tqdm.write so log lines don't
|
|
# break a live progress bar (e.g. the fit bar).
|
|
logger.remove()
|
|
logger.add(lambda m: tqdm.write(m, end=""), colorize=True,
|
|
format="<level>{level.icon}</level> {message}", level="INFO")
|
|
for _lvl, _icon in (("INFO", "I"), ("WARNING", "W"), ("ERROR", "E"), ("DEBUG", "D")):
|
|
logger.level(_lvl, icon=_icon)
|
|
|
|
ROOT = Path(__file__).resolve().parent
|
|
ART = ROOT / "artifacts"
|
|
DEVICE = "cuda"
|
|
DTYPE = torch.bfloat16
|
|
|
|
# The authors publish pre-fitted Jacobian lenses on the Hub (raw Salesforce-
|
|
# wikitext, n=1000 where the _n1000 suffix is present). Loading one beats fitting
|
|
# locally: identical estimator, 1000 prompts, zero compute. Keyed by HF model id;
|
|
# see github.com/anthropics/jacobian-lens walkthrough.ipynb.
|
|
LENS_REPO = "neuronpedia/jacobian-lens"
|
|
LENS_REVISION = "qwen-n1000"
|
|
HUB_LENS_FILE = {
|
|
"Qwen/Qwen3.5-4B": "qwen3.5-4b/jlens/Salesforce-wikitext/Qwen3.5-4B_jacobian_lens_n1000.pt",
|
|
"Qwen/Qwen3.6-27B": "qwen3.6-27b/jlens/Salesforce-wikitext/Qwen3.6-27B_jacobian_lens_n1000.pt",
|
|
"Qwen/Qwen3-4B": "qwen3-4b/jlens/Salesforce-wikitext/Qwen3-4B_jacobian_lens.pt",
|
|
"Qwen/Qwen3-8B": "qwen3-8b/jlens/Salesforce-wikitext/Qwen3-8B_jacobian_lens.pt",
|
|
"Qwen/Qwen3-14B": "qwen3-14b/jlens/Salesforce-wikitext/Qwen3-14B_jacobian_lens.pt",
|
|
"Qwen/Qwen3-32B": "qwen3-32b/jlens/Salesforce-wikitext/Qwen3-32B_jacobian_lens.pt",
|
|
}
|
|
|
|
|
|
def hub_lens_file(model_name: str) -> str:
|
|
"""Filename of the authors' pre-fitted lens for `model_name` inside LENS_REPO.
|
|
KeyError (fail fast) if they don't publish one -- then fit locally via fit.py."""
|
|
return HUB_LENS_FILE[model_name]
|
|
|
|
|
|
def chat_corpus(tok, n_prompts: int) -> list[str]:
|
|
"""jlens's WikiText prompts wrapped in the chat template, for the LOCAL-FIT
|
|
fallback only (the default demo path loads the authors' pre-fitted RAW-wikitext
|
|
lens). Hypothesis: fitting on chat-formatted text puts J closer to the
|
|
distribution we steer in; run-524's verified vectors were fit this way, but it
|
|
was never compared head-to-head with a raw fit, so treat chat-vs-raw as
|
|
unresolved. Called via a lambda in fit_cached, so WikiText only downloads 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}"
|