Hy-MT2-1.8B-JP-Manga-Finetune-v4

This repo contains a version of the Hy-MT2-1.8B model that's been finetuned for Japanese to English translation of conversational/manga-style Japanese. It is the successor to the v2 finetune on the same Japanese → English line and is the recommended model for that pair. Internally this checkpoint is tagged v4.3.

What changed from v2. The training mix now includes Japanese as it actually comes out of an OCR/recogniser pipeline — dropped small kana, merged or split characters, the usual damage — alongside clean text. Measured on held-out manga dialogue and paired against v2, clean-text quality is on par, and quality on damaged recogniser output is better; the gap widens as the input gets noisier. The same training also made the model less likely to invert the direction of an action — who is doing something for, to, or because of whom, in benefactive, passive and causative constructions — which was the most common meaning-level error in v2. That improvement is consistent across the probe but modest; it is not a solved problem, and lines with no named actor remain the hardest.

English only. Like v2, this finetune targets Japanese → English exclusively. It was not trained or evaluated on other target languages, and v2's training of this shape collapsed the base model's non-English output, so expect the same here. For Simplified Chinese, Korean, Vietnamese, European languages and the other targets Hy-MT2 supports, use v3-multilingual — which, on Japanese → English, is behind both v2 and this model.

Scope: this is a manga finetune

The training data is Japanese manga dialogue, and the model is built for translating manga — short, self-contained lines, translated one line at a time.

It will still translate other Japanese, but the finetune's biases follow its training domain, and they are worth knowing before deploying it on something else:

• Fragmentary input gets completed rather than translated. Manga lines are short and elliptical while the training targets are complete, natural sentences, so the model learned to produce a whole utterance from a partial one. Give it truncated input — ASR output, a cut-off caption, a mid-sentence chunk — and it will confidently finish the thought instead of translating only what it was given. Outside the intended domain this is the most common failure.
• Real-world proper nouns are weaker than in the base model. The names in the training data are fictional character names. For public figures, places and brands, stock Hy-MT2-1.8B is often more reliable.
• Technical, business and other specialist terminology is not covered. Nothing in the training data teaches it.
• Register leans conversational. Formal or expository source text comes out more casual than the source warrants.
• A line is translated on its own. The model does not see the surrounding page, so a pronoun or subject that Japanese leaves implicit is resolved from the line itself and general priors. Where the line genuinely cannot say who the speaker means, it will still commit to a reading rather than hedge.

If your input is continuous speech or long-form prose rather than discrete lines of dialogue, A/B it against stock Hy-MT2-1.8B before committing — this finetune does not dominate the base model everywhere.

Prompting

The finetune was trained with a short terminology block in the prompt: a few reference translations, written in English, placed ahead of the Japanese text to translate. It does attend to that block — ablating it on the v2 finetune measurably lowered quality, with some conspicuous individual failures. This is a real difference from stock Hy-MT2-1.8B, which largely ignores terminology injected into the prompt. If you have a glossary, putting it in the prompt is worthwhile here.

The instruction frame is English and the base model's own chat template applies.

Sampling

Recommended: temperature 0.15, top_k 20, top_p 0.6, repeat_penalty 1.05, min_p 0, with the sampler chain ordered penalties → top_k → top_p → temperature.

Higher temperatures measured worse. The cost is mostly variance rather than average quality, and that variance shows up as occasional invented content — so if you are seeing hallucination, lower the temperature before anything else. Note that llama.cpp and llama-server default min_p to 0.05; set it to 0 to match the conditions this model was evaluated under.