Base and Instruct
Pretraining, instruction tuning, long-context adaptation, and Taiwan-local benchmark gates.
Training
Train models with auditable recipes.
OpenFormosa focuses on tokenizer design, pretraining, fine-tuning, evaluation gates, and release documentation. Training inputs must already be licensed, public, synthetic, or separately approved before they appear in this project.
Model training
OpenFormosa-Base
The shared backbone should make Taiwan language, documents, pronunciation notation, and local evaluation behavior visible in one inspectable training path.
- Tokenizer first. Track Taiwan Mandarin, Bopomofo, Tailo, Taigi/Hakka traces, addresses, forms, and public-sector wording.
- Training recipes. Keep mixture notes, hyperparameters, checkpoints, eval gates, and reproducibility commands close to each release.
- No intake pipeline. This site documents training and evaluation; it is not a place to submit or transfer datasets.
Training tracks
Each model branch should have a small, inspectable path from recipe to checkpoint to benchmark result.
ASR correction
Adapters and correction models for transcripts, timestamps, diarization hints, and Taiwan-local terminology.
TTS normalization
Pronunciation, text normalization, style control, and release checks before any generated voice demo.
OCR extraction
Layout parsing, table extraction, address handling, confidence reporting, and privacy review for document tasks.
Release gates
A checkpoint is not a release.
Before publishing weights or demos, each run needs a model card, training-data summary, benchmark report, limitation notes, and acceptable-use boundary.