mirror of
https://github.com/wassname/steer-heal-love.git
synced 2026-06-27 16:47:16 +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>
77 lines
3.1 KiB
Python
77 lines
3.1 KiB
Python
"""Q1 trait-persistence: does the trained adapter move the profile AWAY from base,
|
|
or is it a coherent no-op (healed == reverted to base)?
|
|
|
|
adapter_ppl < steered_ppl only says the adapter is coherent. A do-nothing adapter
|
|
is also coherent. The distinguishing check is the profile DELTA vs base: if
|
|
heal kept the trait, socialnorms/care shift in the steering direction while
|
|
coherence holds. If base == adapter, the adapter learned nothing.
|
|
|
|
Run: uv run python scripts/diag_heal.py out/<ts>_<slug>/ckpt/r0.safetensors
|
|
"""
|
|
|
|
import sys
|
|
|
|
import torch
|
|
import tinymfv
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
sys.path.insert(0, "src")
|
|
from steer_heal.config import RunConfig # noqa: E402
|
|
from steer_heal.steering import teacher_vec # noqa: E402
|
|
from steer_heal.ws.bake import AdapterSpec, baked # noqa: E402
|
|
|
|
ckpt = sys.argv[1]
|
|
N_VIG = None if (len(sys.argv) > 2 and sys.argv[2] == "all") else int(sys.argv[2]) if len(sys.argv) > 2 else 24
|
|
NO_STEER = "--no-steer" in sys.argv
|
|
cfg = RunConfig(n_prompts=12)
|
|
MODEL = cfg.model
|
|
|
|
tok = AutoTokenizer.from_pretrained(MODEL)
|
|
if tok.pad_token is None:
|
|
tok.pad_token = tok.eos_token
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
MODEL, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="eager"
|
|
).eval()
|
|
|
|
|
|
def profile(label):
|
|
rep = tinymfv.evaluate(model, tok, name="classic", n_vignettes=N_VIG,
|
|
conditions=("other_violate",), max_think_tokens=cfg.eval_think_tokens, device=model.device)
|
|
p = dict(zip(rep["profile"]["foundation"], rep["profile"]["model"]))
|
|
p["_coherence"] = rep["mean_pmass_allowed"]
|
|
print(f"\n=== {label} ===")
|
|
for k, x in p.items():
|
|
print(f" {k:12s} {x:.4f}")
|
|
return p
|
|
|
|
|
|
# three points: base, in-band steered (the raw teacher), trained adapter.
|
|
print(f"n_vignettes={N_VIG} no_steer={NO_STEER} ckpt={ckpt}")
|
|
base = profile("BASE (no adapter)")
|
|
|
|
if NO_STEER:
|
|
steer = base # skip the slow steered eval; deltas vs steer are then 0
|
|
else:
|
|
v = teacher_vec(model, tok, cfg)
|
|
with v(model, C=v.cfg.coeff):
|
|
steer = profile(f"STEERED (raw, c={v.cfg.coeff:.1f})")
|
|
|
|
spec = AdapterSpec.from_checkpoint(model, ckpt)
|
|
with baked(model, [spec]):
|
|
adapt = profile(f"ADAPTER (r0, baked) {ckpt.split('/')[-1]}")
|
|
|
|
# SHOULD: adapter delta has the SAME SIGN as steered delta on the trait axis
|
|
# (socialnorms, care) -> heal kept the trait. If adapter delta ~ 0 -> no-op
|
|
# (we "healed" by reverting to base). Coherence: steered may drop, adapter holds.
|
|
print("\n=== trait axis: did the adapter keep the steering direction? ===")
|
|
print(f" {'foundation':12s} {'base':>8s} {'steer':>8s} {'adapt':>8s} "
|
|
f"{'d_steer':>8s} {'d_adapt':>8s} same_sign")
|
|
keys = [k for k in base if not k.startswith("_")]
|
|
for k in sorted(keys, key=lambda k: -abs(steer[k] - base[k])):
|
|
ds, da = steer[k] - base[k], adapt[k] - base[k]
|
|
same = "YES" if (ds * da > 0 and abs(da) > 0.01) else ("no-op" if abs(da) < 0.01 else "OPPOSITE")
|
|
print(f" {k:12s} {base[k]:+8.3f} {steer[k]:+8.3f} {adapt[k]:+8.3f} "
|
|
f"{ds:+8.3f} {da:+8.3f} {same}")
|
|
print(f" {'coherence':12s} {base['_coherence']:8.3f} {steer['_coherence']:8.3f} "
|
|
f"{adapt['_coherence']:8.3f}")
|