mirror of
https://github.com/wassname/jsteer.git
synced 2026-10-17 10:30:22 +08:00
wassnameandClaudypoo
160d675c61
The old rubric pmass was softmax mass on digit tokens at the hand-fed `{"ans": `
slot, so it was ~always 1 (the prefix forces a digit even from a fried model) --
a blind coherence guard. Replace with the users design: free-generate
{"ans":N,"why":str,"2+2":M}, gate coherence on valid-JSON AND 2+2==4. GPU UATs:
object stays valid at |C|<=0.5, breaks at |C|>=0.75 (task 22/23); the old slot
never caught this.
span_pmass (mean top-1 prob over the span) is kept as a within-coherent confidence
read but NOT the coherence signal: a steer-fried model collapses into a confident
degenerate loop, so span_pmass climbs back to ~0.97 while the object is broken
(C=3.0). plot_sweep therefore colors by valid_frac, which cannot be fooled by
confident garbage.
Also: show_steers per-C cowsay now speaks steer-PROMOTED tokens (top of
steered-baseline logits) instead of lens_topk at the last chat-prompt position,
which only ever surfaced think-openers (Okay/Here/The) for every C. UAT: joy steer
-> < joy . happy . Happy . happy > at C=0.3 (task 24). compute_slice stays the
calibrated cross-layer lens readout.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
45 lines
2.1 KiB
Python
45 lines
2.1 KiB
Python
"""Demo artifact: coherence_sweep that WALKS PAST the coherence edge, so the table +
|
|
plot show the full dose-response (ans rises with C) colored by coherence (valid-object
|
|
fraction) with a red-edged incoherent boundary. (Claude) step=0.25 reaches the ~C=1.0
|
|
break found in uat_coherence_break; n_samples=3 for BMA error bars.
|
|
|
|
uv run python scripts/scratch/sweep_artifact.py
|
|
"""
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
|
|
import config # noqa: E402
|
|
import torch # noqa: E402
|
|
from loguru import logger # noqa: E402
|
|
from tabulate import tabulate # noqa: E402
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer # noqa: E402
|
|
|
|
from jsteer import Jacobian # noqa: E402
|
|
from jsteer.demo import coherence_sweep, plot_sweep # noqa: E402
|
|
|
|
MODEL = "Qwen/Qwen3.5-4B"
|
|
OUT = "/tmp/claude-1000/sweep_json_coherence.png"
|
|
tok = AutoTokenizer.from_pretrained(MODEL)
|
|
model = AutoModelForCausalLM.from_pretrained(MODEL, dtype=torch.bfloat16).to("cuda").eval()
|
|
jac = Jacobian.from_pretrained(config.LENS_REPO, filename=config.hub_lens_file(MODEL),
|
|
revision=config.LENS_REVISION)
|
|
band = jac.steer_band(model)
|
|
v = jac.word_vector(model, tok, ["happy", "joy"], layers=band)
|
|
|
|
RUBRIC = "On a scale of 0 (bleak) to 9 (joyful), how is your outlook on the week?"
|
|
rows = coherence_sweep(model, tok, v, RUBRIC, step=0.25, max_steps=5, n_samples=3,
|
|
max_new_tokens=384)
|
|
logger.info("\n" + tabulate(rows, headers="keys", tablefmt="github", floatfmt="+.2f"))
|
|
|
|
fig = plot_sweep(rows, title="joy steer: rubric ans vs C (colored by coherence)")
|
|
fig.savefig(OUT, dpi=110, bbox_inches="tight")
|
|
logger.info(f"wrote {OUT}")
|
|
|
|
# UAT: the sweep must contain BOTH coherent (valid_frac=1) and incoherent (red-edge)
|
|
# rows, so the plot shows the dose-response AND the breakdown edge.
|
|
coh = [r for r in rows if r["coherent"]]
|
|
inc = [r for r in rows if not r["coherent"]]
|
|
logger.info(f"\nUAT: coherent rows={len(coh)}, incoherent (edge) rows={len(inc)} at "
|
|
f"C={[r['C'] for r in inc]} (SHOULD have >=1 of each -> plot shows the edge)")
|