Sublarr/anime-translator-v6-GGUF

anime-translator-v6 (Beta)

This model is in Beta. It was trained and benchmarked on a single show (Jujutsu Kaisen S01E01 vs. Crunchyroll DE ground truth). Results may vary on other series. Feedback and corrections welcome — open an issue or reach out via Sublarr.

A Gemma-3-12B fine-tuned EN→DE subtitle translation model specialized for anime dialogue. Designed to run locally via Ollama, integrated as a community backend in Sublarr.

Why this model?

General-purpose LLMs add meta-commentary ("Here is the translation:") that breaks subtitle formatting. Dedicated document translators (e.g. Hunyuan-MT-7B) hallucinate on short 2–5 word anime lines. This model is fine-tuned specifically on anime subtitle pairs — clean, concise output at 7 GB.

Benchmark (JJK S01E01 vs. Crunchyroll DE, 30 aligned pairs)

Model	BLEU-1	BLEU-2	Length-Ratio	Issues
anime-translator-v6	0.281	0.111	1.02	0 meta-commentary
qwen2.5:14b	0.264	0.125	0.96	Adds "Hier ist die Übersetzung:"
hunyuan-mt-7b	0.141	—	12.76	Verbose, hallucinations on short lines

V6 wins on BLEU-1 (most relevant for subtitle length) and produces the cleanest output.

Training Details

Parameter	Value
Base model	google/gemma-3-12b-it
Training data	OPUS OpenSubtitles v2018 EN-DE (~75k filtered pairs)
LoRA Rank	32 / Alpha 64
Epochs	2
Learning rate	1e-4
Max seq len	256
Hardware	RTX 5090 (32 GB VRAM, ~80 min)
Quantization	Q4_K_M GGUF

Key design decision: Uses apply_chat_template consistently for both training and inference. This format consistency is the core fix over V1–V5 which used seq2seq bases with mismatched formats.

Quick Start with Ollama

# Option 1: Direct pull from HuggingFace (Ollama ≥ 0.3)
ollama pull hf.co/sublarr/anime-translator-v6-GGUF:Q4_K_M

# Option 2: Manual create with Modelfile
git clone https://huggingface.co/sublarr/anime-translator-v6-GGUF
ollama create anime-translator-v6 -f anime-translator-v6-GGUF/Modelfile

Sublarr Integration

This model is available as a one-click community backend in Sublarr:

1. Open Sublarr → Settings → Translation Backends
2. Select "Community Models" → "Anime Translator V6"
3. Click "Install" → Ollama pulls the model automatically
4. Set as active backend → done

# Or set manually in sublarr.env:
SUBLARR_OLLAMA_MODEL=anime-translator-v6

Prompt Format (Gemma-3 Chat Template)

<start_of_turn>user
Translate to German:

{english_subtitle_text}<end_of_turn>
<start_of_turn>model

Inference Parameters

temperature=0.3
top_p=0.9
top_k=40
repeat_penalty=1.1
num_predict=200

Limitations

• Beta quality: Tested on one show, may degrade on other genres/styles
• EN→DE only: Not designed for other language pairs
• Short lines only: Optimized for subtitle-length text (< 20 words), not document translation
• Anime-specific: May underperform on live-action, documentary, or neutral-register content

Roadmap

• Broader test set (5+ shows, different studios)
• Upload LoRA adapter weights separately (for further fine-tuning)
• V7: Explore NLLB-200 fine-tuning as seq2seq alternative

License

MIT — model weights derived from google/gemma-3-12b-it (Gemma Terms of Use apply). Training data: OPUS OpenSubtitles v2018 (CC BY 4.0).