HPEC2026 / exp7_exp8_operator_study /source /README_source_dumps.md
Celsia's picture
Source dumps and full provenance: manifests, LR selections, cost panel, all logs
556aa92 verified
|
Raw History Blame Contribute Delete
1.8 kB

Source error dumps and provenance

source_dumps/

The per-session dumps that results/errdump_all.npz and final/results/errdump_FINAL.npz merge. Kept separately because a merge is a derived artifact: if the merge logic is ever wrong, these are the originals to rebuild from.

file runs tree it was dumped under
errdump_exp7_full.npz 27 (fa_local) 5e93b59
errdump_exp7_nd.npz 18 (fa_local 3-D/4-D) 5e93b59
errdump_exp8_full.npz 27 (fa_sm) 30efd49
errdump_backfill_full.npz 36 (asa/fa 3-D/4-D retrains) 30efd49
errdump_identity.npz 3 (transformer_asa_2d_id) 30efd49

The tree matters: dump_errors gates on run_input_fingerprint, which hashes the source tree and git rev-parse HEAD. A dump can only be reproduced by checking out the commit its runs-dir was declared at. This is also why the original 477-run sweep can never be dumped again — no commit reproduces its tree; all 83 candidates between 5eeb6f9 and 3d94d2b were checked and none matched.

provenance/

Run manifests, learning-rate selections, the cost panel, and the full logs of every sweep, probe, evaluation and dump. exp7_manifest_backup/ is the Exp 7 manifest as originally declared, kept because a resume would have rewritten it.

Two learning-rate selections are here and they are not interchangeable: exp0_mechlocal/lr_selection.json (gru 1e-4, lstm 1e-2, transformer 1e-4 — all three at grid endpoints, so the optimum may lie outside {1e-4 .. 1e-2}) and exp0_mechsm/lr_selection.json (gru 3e-4, lstm 1e-4, transformer 3e-4 — only lstm at an endpoint). Each mixer was given its own Exp 0 arm rather than inheriting the mech rate probed on the authors' operator, because the arms differ by up to 33x on the same backbone.