mirror of
https://github.com/wassname/steer-heal-love.git
synced 2026-06-27 15:17:14 +08:00
wassname
579e1f6671
After verifying guided.py: tinymfv `score` is already a debiased logprob ((lp_fwd+lp_rev)/2, BMA'd), not a "raw logit", and `p = softmax(score)`. My two earlier inventions were both wrong: - log(p) coupled Authority to the other 6 foundations via logsumexp. - the diagonal (auth-blame on auth-vignettes) is pmass-on-correct-label = top1 competence, not the trait, and threw away the FP/FN structure. Use the library-native readout: auth_nats = log(tinymfv profile p[F]) = log of the mean p per foundation over ALL vignettes. For small p, log p ~= logit, so this lands on steering-lite's loading-weighted Δlogit scale (base log(0.099) =-2.3, real shift ~0.5-2 nats). foundation_nats now reads rep["profile"]. Also: - run.py: BLUF `main metric:` line with cue ball (🟢/🟡/🔴 by coherence band). - heal.py: training table to 2 sig figs (nll/kl/loss .2f, gnorm .1f); a per-step loss does not warrant 3 decimals. - diag_stages: accept 1+ ckpts, label each row by its reg from metadata. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
101 lines
4.3 KiB
Python
101 lines
4.3 KiB
Python
"""Target vs off-target effect at each stage, all at the SAME n_vignettes.
|
|
|
|
TARGET = Authority foundation, want DOWN (trait = "do not defer to authority").
|
|
(also report SocialNorms + Care, the axis the 1b note flagged.)
|
|
OFF-TARGET= coherence = tinymfv mean_pmass_allowed = p_any_ans, want HELD ~1.0.
|
|
|
|
Stages: base -> steered(c=0.5,1.0) -> one row per adapter ckpt (labeled by its
|
|
reg). One model load, one vignette set, so every row is paired and comparable.
|
|
|
|
Run: uv run python scripts/diag_stages.py <ckpt1> [ckpt2 ...] [n|all]
|
|
"""
|
|
|
|
import json
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
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.steering import teacher_vec # noqa: E402
|
|
from steer_heal.ws.bake import AdapterSpec, baked # noqa: E402
|
|
|
|
# Trailing "all"/int is the vignette count; everything else is a ckpt path.
|
|
argv = sys.argv[1:]
|
|
N_VIG = None
|
|
if argv and (argv[-1] == "all" or argv[-1].isdigit()):
|
|
N_VIG = None if argv[-1] == "all" else int(argv[-1])
|
|
argv = argv[:-1]
|
|
ckpts = argv # 1+ adapter checkpoints
|
|
|
|
|
|
def ckpt_label(path: str) -> str:
|
|
"""Row label = the run's reg (kl_rev/nll/...) from metadata.json two dirs up."""
|
|
m = json.load(open(Path(path).parents[1] / "metadata.json"))
|
|
reg = m.get("cfg", m).get("reg", "?")
|
|
return f"heal_{reg}"
|
|
|
|
|
|
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()
|
|
|
|
|
|
def prof():
|
|
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)
|
|
p = foundation_nats(rep) # logp per foundation, NATS
|
|
p["coherence"] = rep["mean_pmass_allowed"]
|
|
return p
|
|
|
|
|
|
v = teacher_vec(model, tok, cfg)
|
|
adapters = [(ckpt_label(p), AdapterSpec.from_checkpoint(model, p)) for p in ckpts]
|
|
|
|
rows = {}
|
|
rows["base"] = prof()
|
|
for c in (0.5, 1.0): # 0.5 = coherent operating point; 1.0 = the collapse end
|
|
with v(model, C=c * v.cfg.coeff):
|
|
rows[f"steered(c={c:g})"] = prof()
|
|
for label, spec in adapters:
|
|
with baked(model, [spec]):
|
|
rows[label] = prof()
|
|
|
|
# target = Authority log p (down good, NATS), off-target = coherence (held good).
|
|
# THE Gate-3 question (user): is the trained adapter more coherent PER UNIT behaviour
|
|
# change than raw steering? -> coh_cost = |dCoh| / |dAuth| (coherence lost per nat of
|
|
# Authority shift). LOWER = better pareto. If an adapter has lower coh_cost than the
|
|
# steered rows, distill+heal bought a better behaviour/coherence trade than steering.
|
|
b = rows["base"]
|
|
d_auth_steer = rows["steered(c=0.5)"]["Authority"] - b["Authority"] # retain denom = operating-point shift
|
|
print(f"\nn_vignettes={N_VIG} TARGET=Authority log p (NATS, want DOWN) OFF-TARGET=coherence (want ~{b['coherence']:.2f})")
|
|
print("All foundation columns in NATS (log p, choice-logprob). retain = dAuth(stage)/dAuth(steered c=0.5): "
|
|
"1=heal kept the operating-point trait, 0=reverted to base (UNDO), <0=wrong way.")
|
|
print("coh_cost = |dCoh|/|dAuth| = coherence lost per nat of behaviour change. LOWER is a BETTER pareto. "
|
|
"The point of distill+heal: adapter coh_cost < steered coh_cost. SHOULD: a real HEAL keeps |dAuth| "
|
|
"(retain>0) at near-zero |dCoh| (low coh_cost); an UNDO has retain~0 (no trait, nothing to cost).")
|
|
tbl = []
|
|
for stage, p in rows.items():
|
|
dA = p["Authority"] - b["Authority"]
|
|
dC = p["coherence"] - b["coherence"]
|
|
retain = dA / d_auth_steer if abs(d_auth_steer) > 1e-6 else float("nan")
|
|
coh_cost = abs(dC) / abs(dA) if abs(dA) > 1e-6 else float("nan")
|
|
tbl.append({
|
|
"stage": stage,
|
|
"auth_nats↓": p["Authority"], "dAuth": dA, "retain": retain,
|
|
"socnorm": p["SocialNorms"], "care": p["Care"],
|
|
"coherence→": p["coherence"], "dCoh": dC, "coh_cost↓": coh_cost,
|
|
})
|
|
print(tabulate(tbl, headers="keys", tablefmt="github", floatfmt="+.3f"))
|