Skip to content

ara-diac-1.0

Choose a tag to compare

@ronaldtse ronaldtse released this 25 Aug 03:15
· 653 commits to main since this release
277f51b

ara-diac-1.0

IMF v1 (fp32, decoder kv, opset 14). Trained from rababa train_arabic_r6.py run-006-morph (morphological aux-task: plain + "TAG: "-prefixed iʿrāb stream from qalsadi labels, init from r5); checkpoint rababa-checkpoints:/rababa_arabic_byt5/run-006-morph/best.

field value
task diacritization (Arab → Arab)
artifact ara-diac-1.0-fp32.zip (2.59 GiB)
parts ara-diac-1.0-fp32.zip.part-00 + ara-diac-1.0-fp32.zip.part-01 (GitHub 2GiB cap)
der_total_greedy 2.5793 — greedy decode (the v1 runtime path); SadeedDiac-25, windowed zero-skip at 1400 bytes, 1,200 paragraphs; ByT5-base r6
der_morph_greedy 1.5317 — same harness; Morphological DER (word-final case endings only)
parity cer_delta 0.0pp on 500 samples
sha256 9e6d9777fe2474b150450774aaa1c9e8a9051b502b184fde298b91636dcfdac6
license BSD-3-Clause

Runtimes reassemble split parts transparently and verify every sha256:

from secryst import Model
model = Model.load("ara-diac-1.0")