Kana-Focus/kana-focus-m2m100-418m-ja-en-pt-ct2-int8

Kana Focus M2M100 418M — Japanese to English and Portuguese

This is the offline translation model used by Kana Focus. It adapts Meta AI's MIT-licensed facebook/m2m100_418M model for Japanese learning content and is converted to CTranslate2 INT8 format for local CPU inference.

The model supports:

• Japanese to English
• Japanese to Brazilian Portuguese
• Offline translation after the optional model download
• No account, cloud service, or upload of the learner's text

Training data

The adaptation used 1,972 aligned translation pairs from the Kana Focus sentence library: 986 Japanese-English pairs and 986 Japanese-Portuguese pairs. The split was grouped by Japanese source to prevent the English version of a sentence from leaking into training while its Portuguese version was held out. No OCR capture, private reading material, or unreviewed user text was included.

Evaluation

The final numbers below test the exact CTranslate2 INT8 artifact with beam size 3, not the full-precision training checkpoint.

Target	Model	BLEU	chrF
English	Previous INT8 baseline	27.62	47.81
English	This model	28.25	49.89
Portuguese	Previous INT8 baseline	22.95	44.28
Portuguese	This model	25.79	46.62

Across 196 held-out translations, automated safety checks found no empty outputs, untranslated Japanese residue, or extreme-length outputs. The release manifest contains the exact corpus summary, metrics, warnings, sizes, and SHA-256 hashes.

Runtime files

• model.bin
• config.json
• shared_vocabulary.json
• sentencepiece.bpe.model
• model-release-manifest.json

Limitations

Machine translation can be fluent while still omitting or changing meaning. This model is intended as an educational reading aid, not for legal, medical, financial, safety-critical, or professional translation. Names, idioms, literary language, OCR errors, and uncommon domains can still be mistranslated.

Provenance

Base model: https://huggingface.co/facebook/m2m100_418M

Conversion:

ct2-transformers-converter --model <merged-checkpoint> --quantization int8

See model-release-manifest.json for reproducibility and integrity details.