Models, data & code
Most of what I build is public. Models and corpora live on the Hugging Face Hub; tooling and experiments live on GitHub, much of it under the Twinkle AI community.
TwinkleTokenizer
A 201K byte-level BPE vocabulary trained from scratch on Traditional Chinese — 38% lower tokens/char than Qwen3.8-27B on tw-tokenizer-bench, with a smaller vocabulary.
201,069 vocab · NFC · Apache-2.0 ModelLlama-3.2-Taiwan-3B-Instruct
Instruction-tuned Llama 3.2 adapted to Taiwanese Traditional Chinese and local context.
71 likes on the Hub Model · LegalLlama-3.2-Taiwan-Legal-3B-Instruct
Legal-domain continued pretraining and SFT on Taiwanese law, aimed at statute reasoning and citation.
Token classificationprivacy-filter-tw
PII detection for Traditional Chinese text — the model behind a redaction pipeline for zh-TW corpora.
fineweb-zhtw
A large-scale Traditional Chinese web corpus, filtered and deduplicated for LLM pretraining.
48.1M rows Benchmarktw-tokenizer-bench
Traditional Chinese tokenizer benchmark across five formal-register domains, with fixed character windows and training-contamination shards excluded.
24,294 samples · 18.4M chars Benchmarktw-legal-benchmark-v2
Taiwanese legal reasoning benchmark, widely used to evaluate zh-TW models on statute knowledge.
17k questions · 83 likes Corpustw-news-551M
Traditional Chinese news corpus for domain-adaptive pretraining.
649k documents Dataset · Legalchronolex-tw
Time-aware legal dataset tracking how Taiwanese statutes change across amendments.
42 likesopen-sheet
A spreadsheet framework built for agents — write the model as React, export a live .xlsx.
114 stars Pythonhsun-trainer
End-to-end LLM training for Claude Code: continued pretraining, SFT, and GRPO with Traditional Chinese reward functions.
Python · Drug discoverypeptide-dpt
PSMA-targeted peptide design with a SELFIES-based Gemma model and GRPO docking rewards.
Listawesome-taiwan-project
A curated list of Taiwanese indie developer projects, sites, apps, and open source.
17 stars TypeScripttaiwan-gov-redesigns
Alternative designs for Taiwanese government websites, starting with data.gov.tw.
Notebookdrug_discovery
VAE-based SMILES generation for novel compound design.
Hub
Community hub for Traditional Chinese AI datasets and tooling — hub.twinkleai.tw.
186 stars Translationrlhf-book-zh-tw
Traditional Chinese translation of the RLHF Book, with runnable labs.
177 stars FrameworkEval
Twinkle Eval — a fast LLM evaluation framework built on parallel API calls, covering TMMLU+, tw-legal-benchmark and more.
109 stars Researchllm-lab
Open research lab for Traditional Chinese language model development.
53 stars Translationultrascale-playbook-zh-tw
Traditional Chinese edition of The Ultra-Scale Playbook, with interactive experiments.
45 stars CommunityLLM-Book-Club
Hands-on notebooks and materials for the Twinkle AI book club.
47 stars