Skip to content

Commit 3e4effb

Browse files
kmbandyclaude
andcommitted
ml8 gauntlet: add stage 7 (token-matched content sweep) + stage 8 (deterministic re-measure + seed sweep)
Stage 7: ct_wiki/mix/code/math/chat @ --token-budget 20966 (token-matched to wiki's 21k cap), q3 recipe, --holdout-eval — isolates corpus CONTENT from SIZE. Stage 8: d_q1_off/d_q2_acts/d_q3_s42 clean faithful ladder + d_q3_s0..s4 --rotation-seed best-of-N. Both are the queued payoff runs; stage 8 needs ML8_DETERMINISTIC working at full depth first (blocked on the layer-3 attention NaN under use_deterministic_algorithms — KG 0d841335). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 parent 83824f4 commit 3e4effb

1 file changed

Lines changed: 23 additions & 0 deletions

File tree

scripts/calibration/method_gauntlet.py

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -123,6 +123,29 @@
123123
("ct_chat", {"--corpus": "chat", "--token-budget": "20966", "--seq-len": "2048",
124124
"--faithful-acts": True, "--faithful-weights": True}),
125125
],
126+
8: [ # DETERMINISTIC re-measurement + best-of-N seed search. Calibration is now
127+
# bit-reproducible (calibrate_ml8_paged determinism, default-on), so the run-to-run
128+
# noise floor is ZERO — the faithful ladder deltas below are now trustworthy signal,
129+
# and --rotation-seed is a real reproducible lever. Run with --holdout-eval: SELECT the
130+
# best seed on the never-train held-out (quant_so), REPORT on wiki.test (independent).
131+
# Clean faithful ladder (same recipe as the historical q1/q2/q3, now deterministic):
132+
("d_q1_off", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki"}),
133+
("d_q2_acts", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
134+
"--faithful-acts": True}),
135+
("d_q3_s42", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
136+
"--faithful-acts": True, "--faithful-weights": True}), # rotation-seed 42 (default)
137+
# best-of-N: the q3 recipe, varying ONLY the rotation seed → the calibration search surface.
138+
("d_q3_s0", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
139+
"--faithful-acts": True, "--faithful-weights": True, "--rotation-seed": "0"}),
140+
("d_q3_s1", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
141+
"--faithful-acts": True, "--faithful-weights": True, "--rotation-seed": "1"}),
142+
("d_q3_s2", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
143+
"--faithful-acts": True, "--faithful-weights": True, "--rotation-seed": "2"}),
144+
("d_q3_s3", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
145+
"--faithful-acts": True, "--faithful-weights": True, "--rotation-seed": "3"}),
146+
("d_q3_s4", {"--n-samples": "128", "--seq-len": "2048", "--corpus": "wiki",
147+
"--faithful-acts": True, "--faithful-weights": True, "--rotation-seed": "4"}),
148+
],
126149
}
127150

128151

0 commit comments

Comments
 (0)