kevbart/sticker-studio-models

Sticker Studio models

Models used by Sticker Studio (a personal sticker / pen-plotter app, github.com/kevbart/Misc_Programs, SISER/app), downloaded by its setup so every computer gets the same files. Each file's SHA-256 is in SHA256SUMS.txt and pinned by the app.

mongolian/ - Cyrillic Mongolian word -> traditional Mongolian script (fallback converter)

• byt5-mn/: google/byt5-small (Apache-2.0) fine-tuned on Cyrillic -> traditional word pairs: Greg Eck and Mazaalai lists (CC0, suragch/mongolian-datasets), Wiktionary forms (CC BY-SA), and gege-converter's lexicon tiers (gege-converter, MIT). One target per word, verified sources first; 2 % of words held out (heldout.json). Training script: train_mn_fallback.py.
• byt5-mn-int8/: the same model converted to CTranslate2 int8 (run with compute_type int16 and length_penalty 0).
• Measured (2026-10-14, on the owner's PC): 74.3 % of held-out words exact; as the fallback inside the app's combined converter (gege + verified overrides + spelling rules), MiLiC-Eval traditional script chrF 78.5, CER 9.22 %. Only for words the converter cannot resolve.

inner-mongolian/ - LMT-60 converted to CTranslate2 int8

• NiuTrans/LMT-60-4B and NiuTrans/LMT-60-1.7B (Apache-2.0, NiuTrans), converted unchanged to CTranslate2 int8.
• Measured: FLORES-mvf (Inner Mongolian, traditional script) chrF 49.3 for the 4B int8 conversion against 49.5 for the original weights.

Licences: model weights as their upstream (Apache-2.0); the ByT5 fine-tune's training data includes CC BY-SA Wiktionary material - attribution: Wiktionary contributors, Greg Eck, Mazaalai, gege-converter authors.