mirror of
https://github.com/wassname/jsteer.git
synced 2026-10-17 22:20:17 +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>
46 lines
2.2 KiB
Python
46 lines
2.2 KiB
Python
"""UAT: does the JSON-object coherence probe actually CATCH incoherence? (Claude)
|
|
smoke_sweep only reached |C|=0.3 (still coherent). Here we spot-check rubric_score at
|
|
increasing |C| (one seed, greedy) to find where the object breaks: valid/chk_ok should
|
|
flip to False and span_pmass should collapse as steering fries the model. If they never
|
|
do, the probe can't discriminate -> the coherence guard is still blind.
|
|
|
|
uv run python scripts/scratch/uat_coherence_break.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 rubric_score # noqa: E402
|
|
|
|
MODEL = "Qwen/Qwen3.5-4B"
|
|
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 = []
|
|
for C in (0.0, 0.5, 1.0, 1.5, 2.0, 3.0):
|
|
with v(model, C=C):
|
|
e, c = rubric_score(model, tok, RUBRIC, max_new_tokens=384, seed=0)
|
|
rows.append({"C": C, "ans": e, "span_pmass": c["span_pmass"],
|
|
"valid": c["valid"], "chk_ok": c["chk_ok"]})
|
|
logger.info(f"C={C:+.1f} ans={e:+.2f} span_pmass={c['span_pmass']:+.2f} "
|
|
f"valid={c['valid']} 2+2ok={c['chk_ok']}")
|
|
|
|
logger.info("\n" + tabulate(rows, headers="keys", tablefmt="github", floatfmt="+.2f"))
|
|
broke = [r for r in rows if not (r["valid"] and r["chk_ok"])]
|
|
span_range = max(r["span_pmass"] for r in rows) - min(r["span_pmass"] for r in rows)
|
|
logger.info(f"\nUAT: rows where object broke (invalid or 2+2 wrong) = {len(broke)} at "
|
|
f"C={[r['C'] for r in broke]}; span_pmass range={span_range:+.2f} "
|
|
f"(SHOULD: >=1 break at high |C| AND span_pmass falls, else probe is blind)")
|