Italian-extended MTP draft vocabulary for Qwen3.8-Flash-Next (MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark), for issue #86.
- 65,536 ids. SHA-256
3ebdbad493dc265ab028afcef116cc78a19fb912068e8e21c9cc6d44e120a3e5. - Built with
files/build_draft_vocab_extend.py(upstream, AGPL-3.0, commit6b50864):--base draft_vocab_en_code_47k.txt --size 65536 --model <tokenizer dir>. - Corpus:
wikimedia/wikipediaconfig20231101.it, 700 MiB of plain text (Italian Wikipedia, CC BY-SA). No other data. - Floor 47,172 kept whole, 409 ids pinned, 18,363 added by frequency, 400 byte-fallback ids preserved.
- Use:
MTP_DRAFT_VOCAB=/path/to/draft_vocab_it_en_code_65k.txt.