Hy-MT2-1.8B-JP-Manga-Finetune-v5

Hy-MT2-1.8B fine-tuned for Japanese → English translation of manga dialogue, and it seems like this is the best finetune I've made so far. It succeeds v4 and is the recommended model for that pair. This repo carries both the Q4_K_M GGUF and the merged bf16 weights. If you have an Android phone, you can see for yourself how well this fine-tune works for translating manga on-device with the FumetoReaderPlus app.

What changed from v4. v5 was trained on a rebuilt manga corpus with cleaner source text and better reference translations, plus a small set of targeted lines for short names and short words, which v4 tended to misread as sound effects.

Measured on held-out manga. A blind pairwise LLM judge (page context shown, no reference) prefers v5 over v2 on 47.8% of lines vs 36.8%, the rest ties, and counts fewer meaning errors for v5. chrF++ against reference translations ties v2 and v4, and COMET-22 is about one point lower; v5's output is ~7% longer and worded differently from those references, so read the automatic scores as "not worse by wording".

English only. For other target languages use v3-multilingual, which is behind this model on Japanese → English.

Scope

Built for manga: short, self-contained lines translated one at a time, in the register of the source. Its biases follow that domain:

• Fragments get completed, not translated. Truncated input (ASR, a cut-off caption) is finished into a whole utterance. Outside manga this is the most common failure.
• Real-world proper nouns and specialist terminology are weaker than in the base model. Fictional character names are covered; public figures, places, brands and technical vocabulary are not.
• One line at a time. The model does not see the page; where Japanese leaves the speaker open it usually commits to a reading.
• Output runs ~7% longer than v4's, which matters in tight balloons.

Prompting

Same prompt as v2 and v4: an English terminology block, the instruction, then the line, in the base model's chat template. The model attends to the block, so a glossary in the prompt is worthwhile.

Reference the following manga translations:
部室 translates to clubroom
ついたー translates to I'm here!

Translate the following text from Japanese into English as natural, concise manga dialogue. Preserve specific nouns, exact meaning, speaker intent, tone, punctuation, and sound effects. Output only the translated result without any explanation:

<Japanese line>

Sampling

temperature 0.15, top_k 20, top_p 0.6, repeat_penalty 1.05, min_p 0, sampler chain penalties → top_k → top_p → temperature. llama.cpp defaults min_p to 0.05; set it to 0. Higher temperatures measured worse, mostly as occasional invented content.

Files

• manga-v5-Q4_K_M.gguf — 1,133,080,512 bytes, sha256 e4ed3cf8535c4e99e4b9f3d930ecc6f93bc3b3e0b4b5edca872c4cff5a112abe. eos_token_id = 120020 (some converters write 3 for this architecture, and generation then never stops). Same size and speed as v2 and v4.
• model.safetensors with config.json, the tokenizer files and chat_template.jinja — merged bf16 weights, HunYuanDenseV1ForCausalLM, loadable with transformers and trust_remote_code=True like the base model. If you convert to GGUF yourself, check eos_token_id is 120020, not 3.