diff --git a/models.yaml b/models.yaml index 8b397703..2cfad307 100644 --- a/models.yaml +++ b/models.yaml @@ -12,12 +12,15 @@ models: protocol: phoneme-token PER, clean v1 test split (3,210 rows), single-row greedy source: interscript/interscript-train run-080-thai-lex-v6a eval (run-080-thai-lex-v6a/eval-verdict.json lineage; WO67) parity: - samples: 0 + samples: 600 + cer_delta: 0.0001 + row_exact_match: 0.9917 note: >- - PENDING — intermediate comparison exact (encoder 0.000000, step-1 logits - 0.0000, greedy ids identical); the legacy harness contradicts the - proven-exact artifact (CER 1.02 at BS=1) and is under investigation — - gate NOT claimed. + torch vs shipped zip, single-row greedy, 600 clean-split rows + (run-080-thai-v6a-parity-3, 2026-10-24). The two earlier red runs + tested a mislabeled wrong-model export (the R8_RUN env trap); + the corrected artifact is proven exact — encoder hidden 0.000000, + step-1 logits 0.0000 (run-080-thai-parity-dbg-2). license: BSD-3-Clause trained_from: >- ByT5-small; corpora/tha-phonemic-v5.yaml (92,669 length-filtered