AbdallahNasir/gender-mt-gemma2-2b-dpo-fewshot
The AbdallahNasir/gender-mt-gemma2-2b-dpo-fewshot model is a 2.6 billion parameter language model based on the Gemma2 architecture, fine-tuned using Direct Preference Optimization (DPO). It specializes in gender-specific machine translation, specifically trained on TED gender pairs with a 3-shot prompting strategy. This model is designed for tasks requiring nuanced understanding and generation of gendered language, offering an 8192-token context length.
Loading preview...
Overview
The gender-mt-gemma2-2b-dpo-fewshot model is a specialized language model, built upon the Gemma2 architecture with 2.6 billion parameters. It has been fine-tuned using Direct Preference Optimization (DPO), a method known for aligning models with human preferences. The training specifically focused on gender-specific machine translation tasks.
Key Capabilities
- Gender-aware Translation: Optimized for handling and generating gendered language, particularly within the context of machine translation.
- DPO Fine-tuning: Leverages Direct Preference Optimization on TED gender pairs, enhancing its ability to produce preferred outputs.
- Few-shot Learning: Utilizes a 3-shot prompting strategy during training, incorporating demonstrations for male speakers, female speakers, and neutral prompts to guide its understanding.
- Gemma2 Architecture: Benefits from the foundational capabilities of the Gemma2 model family.
Good for
- Research and development in gender bias mitigation in NLP.
- Applications requiring nuanced gender representation in text generation or translation.
- Experiments with Direct Preference Optimization on specific linguistic attributes.
- Tasks involving machine translation of content with gendered language, where specific gendered outputs are desired or need to be controlled.