mirror of
https://github.com/wassname/jsteer.git
synced 2026-09-04 01:00:32 +08:00
wassnameandClaudypoo
220f5018ac
Adds this session's evidence: uat_prefitted_4b.py (proves the n1000 lens steers, word>random), calib_c_prefitted.py + calib_persona.py (how the demo Cs were chosen: word knee ~0.5, mean_diff ~1), run_nb.py (nbclient notebook executor, bypasses the broken global nbconvert config). Removes u4_step3_guard.sh / retry.sh; fit.py gains --out for scratch fits to non-canonical paths. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
47 lines
2.0 KiB
Python
47 lines
2.0 KiB
Python
"""UAT: does the authors' pre-fitted n=1000 Qwen3.5-4B lens drive our steering? (Claude)
|
|
|
|
Loads the Hub lens (raw Salesforce-wikitext, n=1000) through Jacobian.load, extracts a
|
|
happy/joy word_vector on the mid-depth band (to match run-524's regime, since the
|
|
pre-fitted lens spans ALL layers 0..30), and shows baseline vs +C vs the random control.
|
|
If +C reads happier and stays coherent while random at the same C does not, the
|
|
pre-fitted raw lens is a drop-in for the demo and local fitting is unnecessary here.
|
|
"""
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
|
|
|
|
import torch
|
|
from huggingface_hub import hf_hub_download
|
|
from loguru import logger
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
import config # noqa: F401 loguru-on-import
|
|
from jsteer import Jacobian, show_steer
|
|
|
|
MODEL = "Qwen/Qwen3.5-4B"
|
|
tok = AutoTokenizer.from_pretrained(MODEL)
|
|
model = AutoModelForCausalLM.from_pretrained(MODEL, dtype=torch.bfloat16).to("cuda").eval()
|
|
|
|
path = hf_hub_download(
|
|
"neuronpedia/jacobian-lens", revision="qwen-n1000",
|
|
filename="qwen3.5-4b/jlens/Salesforce-wikitext/Qwen3.5-4B_jacobian_lens_n1000.pt")
|
|
jac = Jacobian.load(path)
|
|
logger.info(f"pre-fitted lens: {jac!r}")
|
|
|
|
# All-layer lens -> restrict steering to the 0.3-0.9 depth band (n_layers=32), the
|
|
# regime run-524 used; steering all 31 layers at once over-drives the residual.
|
|
n_layers = model.config.num_hidden_layers
|
|
band = [l for l in jac.layers if 0.3 <= l / n_layers <= 0.9]
|
|
logger.info(f"steer band (0.3-0.9 of {n_layers}): {band}")
|
|
|
|
v = jac.word_vector(model, tok, ["happy", "joy"], layers=band)
|
|
v_rand = jac.random_vector(seed=0, layers=band)
|
|
|
|
logger.info("=== WORD vector (happy/joy) on pre-fitted n1000 lens ===")
|
|
show_steer(jac, model, tok, v, "Describe how your week has been going.",
|
|
Cs=(0, 6), max_new_tokens=80)
|
|
logger.info("=== RANDOM control at matched C ===")
|
|
show_steer(jac, model, tok, v_rand, "Describe how your week has been going.",
|
|
Cs=(0, 6), max_new_tokens=80)
|