Downloads · 30 days
10
20% of all-time downloads
Kana-Focus/kana-focus-m2m100-418m-ja-en-pt-ct2-int8
kana-focus-m2m100-418m-ja-en-pt-ct2-int8 is a translation model from Kana-Focus. Use it when you need text moved from one language to another. It is set up for ctranslate2. The card lists the license as mit.
This is the offline translation model used by Kana Focus. It adapts Meta AI's MIT-licensed facebook/m2m100418M model for Japanese learning content and is converted to CTranslate2 INT8 format for local CPU inference.
Downloads · 30 days
10
20% of all-time downloads
All-time downloads
51
Public
Repo size
493 MB
Likes
0
Public
Click a slice to open those files.
.bin491 MB · 99%
From the Hugging Face model README
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:
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.
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.
model.binconfig.jsonshared_vocabulary.jsonsentencepiece.bpe.modelmodel-release-manifest.jsonMachine 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.
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.