hypaai/Hypa-Gemma4-E2B-v1
Hypa-Gemma4-E2B-v1 is a 5.1 billion parameter LoRA-merged fine-tune of Google DeepMind's Gemma 4 E2B-it model, developed by Hypa Intelligence. This model is specifically optimized for multilingual understanding and tool-aware instruction following, with a unique focus on seventeen languages including English, French, Spanish, and fourteen low-resource Nigerian languages. It excels at translation, language detection, and dictionary-style explanations, making it ideal for applications requiring robust multilingual capabilities, especially for underrepresented languages.
Loading preview...
Hypa-Gemma4-E2B-v1: Multilingual & Tool-Aware Gemma Fine-Tune
Hypa-Gemma4-E2B-v1 is a 5.1 billion parameter model from Hypa Intelligence, fine-tuned from Google DeepMind's Gemma 4 E2B-it. Its core differentiator is its specialized focus on multilingual understanding and tool-aware instruction following, particularly for low-resource and underrepresented languages. The model supports seventeen languages, including English, French, Spanish, and fourteen Nigerian languages, some of which have not been formally represented in large-scale fine-tuning corpora before.
Key Capabilities
- Multilingual Translation: Designed for translation between English/French/Spanish and the fourteen covered low-resource languages.
- Language Detection: Capable of detecting all seventeen supported languages.
- Lexical Explanation: Provides dictionary-style explanations and lexical lookups.
- Instruction Following: Excels at general multilingual instruction-following, inheriting Gemma 4's native chat template and role structure.
- Tool-Aware Prompting: Retains the base model's dedicated formatting for thinking and tool use, enabling agentic prompting structures.
Training & Performance Notes
The model was trained using LoRA via Unsloth and QLoRA on 15.9 million examples. While the final merged 16-bit weights are provided, the developers recommend using the LoRA checkpoint at step 40,000 for optimal generalization, as the final checkpoint showed signs of overfitting. Qualitative observations indicate minor improvements over the base Gemma 4 E2B, with significant gains for the smallest languages. The model is released under the Apache 2.0 License, making it suitable for both research and commercial applications.