docs: make active-path comments concise

This commit is contained in:
wassname
2026-06-10 05:19:52 +00:00
parent c031d9db76
commit 51c5a757ef
5 changed files with 50 additions and 122 deletions
+6 -6
View File
@@ -1,4 +1,4 @@
"""Reproduce a finished run's paired knob-off/knob-on final-test evaluation."""
"""Reproduce a finished run's paired quarantine-ablated/enabled final-test evaluation."""
from __future__ import annotations
import json
@@ -45,8 +45,7 @@ def main(run_dir: Positional[Path]) -> None:
wrappers[name]["delta_S_hack"].data.copy_(delta_hack[name].to(device, torch.bfloat16))
prior_eval = json.loads((run_dir / "deploy_test.json").read_text())
# by_mode keys ARE the modes the original deploy eval spanned (present in every json
# version); reproduce the same set so the re-scored knob-off matches the headline.
# Reproduce the original evaluation environment exactly.
eval_modes = sorted(prior_eval["by_mode"].keys())
_, problems = load_eval_splits(eval_modes, cfg["eval_n_prompts"])
gen_cfg_eval = GenerationConfig(
@@ -56,7 +55,7 @@ def main(run_dir: Positional[Path]) -> None:
)
eval_idxs = list(range(len(problems)))
torch.manual_seed(EVAL_GEN_SEED)
with ablate_quarantine(wrappers): # knob OFF = the deployed model
with ablate_quarantine(wrappers): # quarantine ablated = deployed model
ev = eval_hack_solve(
model, tok, problems, eval_idxs, gen_cfg_eval, device, cfg["max_new"], cfg["eval_batch_size"])
torch.manual_seed(EVAL_GEN_SEED)
@@ -74,8 +73,9 @@ def main(run_dir: Positional[Path]) -> None:
for m, (h, v, s, c) in ev["by_mode"].items()},
}
(run_dir / "deploy_test.json").write_text(json.dumps(out, indent=2))
logger.info(f"FINAL paired test n={ev['n']}: knob-off hack={ev['hack']:.3f} solve={ev['solve']:.3f}; "
f"knob-on hack={ev_on['hack']:.3f} solve={ev_on['solve']:.3f}")
logger.info(f"FINAL paired test n={ev['n']}: quarantine-ablated hack={ev['hack']:.3f} "
f"solve={ev['solve']:.3f}; quarantine-enabled hack={ev_on['hack']:.3f} "
f"solve={ev_on['solve']:.3f}")
for m, d in out["by_mode"].items():
logger.info(f" {m:14s} hack={d['hack']:.3f} vhack={d['vhack']:.3f} solve={d['solve']:.3f} n={d['n']}")