TabView M2M100 418M — CTranslate2 INT8 artifact (v1)

This repository contains a TabView-specific converted artifact. It is not the official upstream Meta model repository. The upstream model is facebook/m2m100_418M.

What is in it

tabview-m2m100-418m-ct2-int8-v1.zip holds one folder, m2m100-418m-ct2-int8-v1/, with exactly:

• model.bin, config.json, shared_vocabulary.json — the CTranslate2 INT8 model
• sentencepiece.bpe.model — the upstream SentencePiece model, unchanged
• manifest.json — provenance and the SHA-256 of every file (also at the root of this repository)
• README.md, LICENSE

The weights are not modified beyond the format conversion and INT8 quantization. The converter loaded them from the Hub's safetensors conversion of that revision (refs/pr/27, ae6407352f7f86e328c0f0678d466fc4246e78b3); its 509 tensors equal the revision's pytorch_model.bin (3 tied copies of the shared embedding are stored once).

Conversion command (separate build environment; torch/transformers are not part of the runtime):

ct2-transformers-converter --model facebook/m2m100_418M --revision 55c2e61bbf05 --output_dir ct2_int8 \
    --quantization int8 --copy_files sentencepiece.bpe.model tokenizer_config.json special_tokens_map.json vocab.json

How TabView uses it

The TabView Operator downloads the archive from a URL pinned to an exact commit of this repository, checks the archive SHA-256 and every file's SHA-256 against values compiled into the program, and only then installs it. Translation runs entirely on the user's PC (CTranslate2 + SentencePiece; source token __xx__, target prefix __ko__) — review text is never sent to a translation service.

Citation

Fan et al., Beyond English-Centric Multilingual Machine Translation, 2020.