17slever17/translate-gemma-4-sub-e4b
17slever17/translate-gemma-4-sub-e4b is a 7.9 billion parameter multilingual Gemma 4 fine-tune, specifically optimized for both general translation and context-aware subtitle translation. This model excels at preserving meaning, tone, and conversational style by utilizing previous source and translated segments as context. It is particularly effective for live and conversational speech, supporting custom style instructions, speaker information, and terminology rules.
Loading preview...
Translate Gemma 4 Sub E4B: Context-Aware Multilingual Translation
Translate Gemma 4 Sub E4B is a 7.9 billion parameter multilingual fine-tune of the Gemma 4 model, developed by 17slever17. It is specifically designed for both general translation and, uniquely, for context-aware subtitle translation. This model is a core component of the open-source SubWave real-time subtitle translator.
Key Capabilities
- Context-Aware Subtitle Translation: Unlike many other translation models, Translate Gemma 4 Sub can leverage previous source and translated subtitle segments as context. This is crucial for accurately translating phrases split across segments or those that lose meaning in isolation.
- Preserves Nuance: Optimized to maintain meaning, tone, slang, uncertainty, repetitions, incomplete speech, and natural conversational style.
- Customizable Translation: Supports custom style instructions, speaker information, terminology rules, and glossaries to guide translations.
- Multilingual Support: Retains the language coverage of the underlying Gemma 4 model, with strong performance in English, Russian, Spanish, German, and Japanese, and good support for French, Portuguese, Chinese, Dutch, Italian, and Korean.
Performance Highlights
Benchmarks show that Translate Gemma 4 Sub E4B outperforms other Gemma 4 variants in several key metrics for translation quality, including chrF++, BERTScore, and COMET-DA, particularly in contextual translation scenarios. It also demonstrates slight improvements in general translation quality over base Gemma 4 models, as evidenced by FLORES-200 evaluations.
Ideal Use Cases
This model is particularly well-suited for:
- Live and Conversational Speech Translation: Excels in scenarios like streaming content where context, speaker intent, and tone are critical.
- Subtitle Translation: Designed to handle the complexities of subtitle translation, including fragmented sentences and conversational nuances.
- Applications Requiring Stylistic Control: Useful when specific translation styles, tones, or adherence to glossaries are needed.