fas-diac-base-1.0

Persian diacritization, base tier (run-077): byt5-small seq2seq on the 445,869-pair HomoRich-derived corpus (CC0-1.0). IMF v1 fp16.

Metrics (9,288-row docs split — byte-identical to the rababa-farsi

reference split; raw char CER, both conventions)

Row-mean CER 0.603% · corpus-level CER 0.777% · word accuracy 82.33%. The historical 0.52% line is corpus-level under run-001's undocumented recipe and does not reproduce (five controlled arms, WO47); this is the best verified number.

Improvement history (all closed with data)

epochs (3→6: 0.65→0.60) · backbone (small/base) · optimizer (docs-exact 5e-4/linear) · metric convention (row-mean vs corpus) · multi-arm oracle distillation (run-080/081: arms too correlated, 10.0% disagreement — MISS). Remaining untried: corpus rebuild to the raw 528,837 rows.

License

Training data: HomoRich-G2P-Persian, CC0-1.0 (verified via HF cardData). Weights and code: BSD-3-Clause.