Legacy v1 rebaseline
Stub README (2026-09-28): this directory predates the README convention. The summary below was reconstructed from its files; see them for detail.
Question. What held-out PPL does the legacy v1 trainer reach on Shakespeare 1M (d=16, 10 SE, anc-grad off) from the shared seed models, as a parity reference for v2?
Answer. Seed 1 reached 9.58 and seed 2 reached 8.89 (legacy sliding-window PPL, d=16, 10k positions of a 50k-char held-out file). The landing commit notes that v2 beat legacy by about 0.9 PPL at this config.
- v1 trainer, 10 SE, anc-grad off, seed 1: legacy sliding-window held-out PPL (d=16, 10k positions) = 9.5843
- v1 trainer, 10 SE, anc-grad off, seed 2: legacy sliding-window held-out PPL (d=16, 10k positions) = 8.8919
Sources. Numbers come from shakespeare/seed{1,2}_off/heldout.log. The purpose comes from the message of commit 2b913be ('fresh legacy baseline runs against /tmp/seed{1,2,3}.model for the parity reference').
Caveats. The commit mentions seeds 1-3, but only seeds 1 and 2 are present. RMSProp lr=3e-3 warmup-cosine, pd=1. The held-out file is 50k tokens (probably the tail of data/input.txt, and possibly inside the training trie (unverified). rnd/TRIAGE.md lists this dir as affected by the 2026-05-26 loss fix.