Overview

ReBatch/Llama-3-8B-dutch is an 8 billion parameter conversational model, fine-tuned from Meta's Llama 3 8B. It was developed by ReBatch using QLORA and ORPO techniques, specifically trained on the BramVanroy/ultra_feedback_dutch synthetic feedback dataset to enhance its Dutch language capabilities.

Key Capabilities

• Dutch Conversational AI: Optimized for generating and understanding Dutch text in chat-based applications.
• Fine-tuned Performance: Utilizes ORPO (Odds Ratio Preference Optimization) for alignment with high-quality synthetic feedback, aiming for improved conversational quality.
• Efficient Training: Trained using QLORA in bfloat16 with Flash Attention 2 on a single H100 80GB GPU.

Evaluation and Limitations

Evaluated using scandeval, the model showed mixed results, with some improvements and some decreases in scores compared to the base Llama 3 8B Instruct model, despite being trained on only 200,000 samples for one epoch. While aligned with GPT-4-turbo output, users should be aware that the model may still produce incorrect, misleading, or potentially offensive content. Further training with more data or epochs might enhance its performance.