mirror of
https://github.com/wassname/steer-heal-love.git
synced 2026-06-27 19:17:00 +08:00
wassname
48814897ef
Headline (gemma-3-4b-it s42, care-over-authority): aggregating the kl_rev barrier by rmse over token positions (not the mean) holds coherence flat at 0.997 across all 8 rounds, where the mean aggregate collapses to 0.62 by r7 (token loops). Mean dilutes the few incoherent positions under the tau gate; rmse is outlier-sensitive and fires on them. Cost is depth (rmse run leashes to base, trait stays shallow); matched control still running. - plot.py: coherence panel -> log-incoherence (1-coh, log axis, down=coherent); map coherence axis matches; red steer kept on the over-pipeline panels only. - heal.py: fix kl_agg=p95 crash (torch.quantile rejects bf16 -> .float()). - run.py: persist per-round adapter gens (adapter_gen) for the outputs table. - config.py: coh_floor early-stop knob. - README: results table (mean vs rmse), trajectory figure, outputs-over-loop appendix (per-round completions as quotes); spec persona corrected to pos-neg. - docs/reviews: kl_agg review, pool saturation test, care-lens plan. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
60 lines
2.6 KiB
Python
60 lines
2.6 KiB
Python
"""Re-render a run's trajectory plot from its events.jsonl, without re-running the loop.
|
|
|
|
Why this exists: the loop imports plot.py at process start, so a plot-code fix never
|
|
reaches an already-running job. events.jsonl persists every stage (base/steered/healed),
|
|
so we can rebuild the `stages` list offline and call the CURRENT write_trajectory.
|
|
|
|
uv run python scripts/plot_run.py --run-dir out/20260604T172126_gemma-3-4b-it_kl_rev_s42/
|
|
|
|
A run from before the `base` event was persisted has no base node; pass it explicitly
|
|
(tyro reads negatives as flags, so use `=`):
|
|
... --base-auth=-2.35 --base-care=-1.30 --base-coh=0.996
|
|
"""
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
import srsly
|
|
import tyro
|
|
|
|
from steer_heal.plot import write_report, write_trajectory
|
|
|
|
|
|
def main(run_dir: Path, base_auth: float | None = None,
|
|
base_care: float | None = None, base_coh: float | None = None):
|
|
events = list(srsly.read_jsonl(run_dir / "events.jsonl"))
|
|
by_stage = lambda s: [e for e in events if e["stage"] == s]
|
|
|
|
base_ev = by_stage("base")
|
|
if base_ev:
|
|
bm = base_ev[0]
|
|
base = {"auth_nats": bm["auth_nats"], "care_nats": bm["care_nats"], "coherence": bm["coherence"]}
|
|
else:
|
|
# older run: base wasn't persisted. require it on the CLI (fail fast, no silent default).
|
|
assert base_auth is not None, (
|
|
f"{run_dir}/events.jsonl has no `base` event (ran before run.py persisted it); "
|
|
"pass --base-auth/--base-care/--base-coh from the run's base eval log line."
|
|
)
|
|
base = {"auth_nats": base_auth, "care_nats": base_care, "coherence": base_coh}
|
|
|
|
stages = [{"round": "-", "stage": "base", "m": base}]
|
|
steered = {e["round"]: e for e in by_stage("steered_eval")}
|
|
healed = {e["round"]: e for e in by_stage("round")}
|
|
for rnd in sorted(healed): # one steered + one healed per completed round, in order
|
|
for src, kind in [(steered, "steered"), (healed, "healed")]:
|
|
e = src[rnd]
|
|
stages.append({"round": rnd, "stage": kind,
|
|
"m": {"auth_nats": e["auth_nats"], "care_nats": e["care_nats"],
|
|
"coherence": e["coherence"]}})
|
|
|
|
png = write_trajectory(run_dir, stages)
|
|
print(f"re-rendered {png} from {len(stages)} stages ({len(healed)} rounds)", file=sys.stderr)
|
|
|
|
gen_rounds = sorted(by_stage("adapter_gen"), key=lambda e: e["round"])
|
|
if gen_rounds: # runs from before adapter_gen was persisted (e.g. #107/#108) have none
|
|
out = write_report(run_dir, gen_rounds)
|
|
print(f"re-rendered {out} from {len(gen_rounds)} rounds of gens", file=sys.stderr)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tyro.cli(main)
|