wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-06-27 16:30:30 +08:00
Code Issues Packages Projects Releases Wiki Activity

broad: TEACHER_RT -> dense pool (was sparse, under-seeds); log: rename table cols train/deploy (drop 'knob')

Browse Source 
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-06-07 22:12:00 +00:00
parent 484305d7b4
commit caa0d09472
2 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
justfile
+1 -1
View File
@@ -150,7 +150,7 @@ queue-dir6 seed='43':
# Lower priority than queue-dir6 (the focused single-seed set runs first). resolve:
# routeV deploy_hack < vanilla at matched solve, significant across the 3 seeds (paired t,
# alpha=0.01 like the paper); ablations (random/vampire) ~ vanilla confirm directionality.
TEACHER_RT := "out/pools/teacher_pool_runtests"
TEACHER_RT := "out/pools/teacher_pool_runtests_dense"
queue-broad:
#!/usr/bin/env bash
set -eu
src/vgrout/train.py
+6 -6
View File
@@ -1984,15 +1984,15 @@ def main(cfg: Config) -> int:
# --- TAIL: argv, the result table, the single objective. The last lines. ---
# solve and hack alone are gameable (tank solve to kill hack, or accept hack to
# lift solve); the deploy gap solve-hack is the one number to maximise. Taken
# from the FINAL DEPLOY eval (knob-off, held-out test) = the shipped model on
# unseen problems. knob-on hack is the train-rollout student rate (different set,
# so its solve cell is "-": there is no knob-on deploy-style solve to pair it).
# from the FINAL DEPLOY eval (quarantine deleted, held-out test) = the shipped
# model on unseen problems. The "train" hack is the train-rollout student rate
# (different set, so its solve cell is "-": no train deploy-style solve to pair).
_dh, _ds, _dn = ev["hack"], ev["solve"], ev["n"]
_deploy_col = f"knob-off (deploy, test n={_dn})"
_deploy_col = f"deploy (test n={_dn})"
print(f"\n\nargv: {' '.join(sys.argv)}\n")
print(tabulate(
[{"measure": "hack ↓", "knob-on (train)": f"{hack_rate_s:.3f}", _deploy_col: f"{_dh:.3f}"},
{"measure": "solve ↑", "knob-on (train)": "-", _deploy_col: f"{_ds:.3f}"}],
[{"measure": "hack ↓", "train": f"{hack_rate_s:.3f}", _deploy_col: f"{_dh:.3f}"},
{"measure": "solve ↑", "train": "-", _deploy_col: f"{_ds:.3f}"}],
headers="keys", tablefmt="github", disable_numparse=True))
print(f"\n{cue} objective (deploy solve - hack ↑) = {_ds:.3f} - {_dh:.3f} = {_ds - _dh:+.3f} "
f"[arm={cfg.arm} seed={cfg.seed}]")