mirror of
https://github.com/wassname/steer-heal-love.git
synced 2026-06-27 15:17:14 +08:00
wassname
4568ddf491
The trait metric was taking the diagonal of tinymfv's raw pre-softmax BMA `score` logit (unnormalised), giving base Authority ~-5 and absurd 8-nat swings, then comparing those to steering-lite's 0.5-2 nat reference -- which is a DIFFERENT metric (loading-weighted Delta-logit of binary p(is-wrong)). Wrong scale, wrong comparison. Fix: auth_nats = mean log p[authority] on authority-defiance vignettes (the NORMALIZED choice logprob, the diagonal of the softmax `p`). Base ~log(0.099) = -2.3, real shifts ~1-3 nats. DRY: evaluate_model now calls foundation_nats. Also: - diag_stages: steer at operating point c=0.5 (c=1 collapses coherence to 0.05), add coh_cost = |dCoh|/|dAuth| (coherence lost per nat of behaviour) to answer "is the adapter a better pareto than raw steering?". - diag_csweep: drop the bogus 0.5-2 steering-lite anchor; SocialNorms co-moving with Authority is expected (both binding foundations), not collapse. - gitignore out/ and results.tsv (experiment outputs, stale schema). - personas docs (steering-lite proper-pair rules), spec Plans B/C/D, journal. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
93 lines
4.3 KiB
Python
93 lines
4.3 KiB
Python
"""GATE 1: does the steering vector move the target WHILE staying coherent?
|
|
|
|
Sweep steering strength c and, at each c, eval the foundation profile AND
|
|
generate one completion, so we judge both the metric and the text. This is the
|
|
gate that must pass before filter/lora gates matter: if no c gives a target shift
|
|
at coherence ~0.95, the vector is the problem, not the heal.
|
|
|
|
Reading the (dAuth, coherence) pareto:
|
|
PASS a c with large -dAuth at coherence >= ~0.95 (knee before collapse)
|
|
too weak -dAuth ~ 0 until coherence cliffs
|
|
too strong -dAuth only appears once coherence < ~0.85 (no knee, bad pareto)
|
|
wrong target |dCare| or |dSocialNorms| > |dAuth| at the same c
|
|
collapse all foundations shrink proportionally + coherence drops (no specificity)
|
|
|
|
Run: uv run python scripts/diag_csweep.py [n|all]
|
|
"""
|
|
|
|
import sys
|
|
|
|
import torch
|
|
import tinymfv
|
|
from tabulate import tabulate
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
sys.path.insert(0, "src")
|
|
from steer_heal.config import RunConfig # noqa: E402
|
|
from steer_heal.eval import foundation_nats # noqa: E402
|
|
from steer_heal.prompts import POOL, chat_prompt # noqa: E402
|
|
from steer_heal.steering import _gen_one, teacher_vec # noqa: E402
|
|
|
|
N_VIG = None if (len(sys.argv) > 1 and sys.argv[1] == "all") else int(sys.argv[1]) if len(sys.argv) > 1 else None
|
|
CS = [0.0, 0.25, 0.5, 0.75, 1.0, 1.5]
|
|
cfg = RunConfig(n_prompts=12)
|
|
|
|
tok = AutoTokenizer.from_pretrained(cfg.model)
|
|
if tok.pad_token is None:
|
|
tok.pad_token = tok.eos_token
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
cfg.model, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="eager"
|
|
).eval()
|
|
|
|
v = teacher_vec(model, tok, cfg)
|
|
demo_prompt = chat_prompt(tok, cfg.gen_system, POOL[0]) # fixed prompt for the qualitative read
|
|
|
|
|
|
def profile():
|
|
rep = tinymfv.evaluate(model, tok, name="classic", n_vignettes=N_VIG,
|
|
conditions=("other_violate",), max_think_tokens=cfg.eval_think_tokens,
|
|
device=model.device, return_per_row=True)
|
|
nats = foundation_nats(rep) # logp per foundation, NATS
|
|
nats["coherence"] = rep["mean_pmass_allowed"]
|
|
return nats
|
|
|
|
|
|
rows, samples = [], []
|
|
for c in CS:
|
|
with v(model, C=c * v.cfg.coeff):
|
|
p = profile()
|
|
gen = _gen_one(model, tok, demo_prompt, cfg)
|
|
rows.append((c, p))
|
|
samples.append((c, gen))
|
|
|
|
b = rows[0][1]
|
|
print(f"\nn_vignettes={N_VIG} c-sweep of the teacher vector (coeff={v.cfg.coeff}) ALL VALUES IN NATS (log p, choice-logprob)")
|
|
print("auth_sep = base - steered Authority log p (POSITIVE = steered attributes authority-defiance "
|
|
"less to authority = correct direction). Scale is tinymfv's diagonal log(p); base auth_nats "
|
|
"~-2.3, a real shift is ~1-3 nats. NOT steering-lite's 0.5-2 p(is-wrong) metric.")
|
|
tbl = []
|
|
for c, p in rows:
|
|
tbl.append({
|
|
"c": c,
|
|
"auth_nats↓": p["Authority"], "auth_sep↑": b["Authority"] - p["Authority"],
|
|
"care_nats": p["Care"], "care_sep": b["Care"] - p["Care"],
|
|
"socnorm_nats": p["SocialNorms"], "coherence→": p["coherence"],
|
|
})
|
|
print(tabulate(tbl, headers="keys", tablefmt="github", floatfmt="+.3f"))
|
|
print("\nPASS = a c with a clear POSITIVE auth_sep (~1-3 nats on this log(p) scale) AND coherence "
|
|
">= ~0.95. If auth_sep only appears once coherence < 0.85 -> bad pareto (vector too imprecise). "
|
|
"If care_sep ~ auth_sep -> broad permissivizing, not surgical (SocialNorms co-moving is OK).")
|
|
print("SHOULD (signal vs collapse): a REAL trait shift REDISTRIBUTES foundation mass -- some DOWN "
|
|
"(Authority/Care/SocialNorms) some UP (Fairness/Sanctity) -- while coherence falls LESS than "
|
|
"the foundations. GENERAL COLLAPSE instead drops every foundation AND coherence by a similar "
|
|
"fraction (mass leaks off the allowed answer tokens, no redistribution). At the c where "
|
|
"Authority drops, check: do Fairness/Sanctity RISE (signal) or does everything including "
|
|
"coherence fall together (collapse)?")
|
|
print("SHOULD (coherence levels): c=0 MUST be ~1.0 (sanity). >=0.95 mild, 0.85-0.95 degraded, "
|
|
"<0.85 broken. A trait shift is only 'free' if it lands at coherence >=0.95.")
|
|
|
|
# qualitative: read whether the steered text is coherent AND anti-authority.
|
|
print(f"\n=== steered generations (prompt: {POOL[0]}) ===")
|
|
for c, gen in samples:
|
|
print(f"\n--- c={c:g} ---\n{gen}")
|