robinsmits/Qwen1.5-7B-Dutch-Chat: A Dutch-Optimized LLM

This model is a 7.7 billion parameter language model, fine-tuned by robinsmits, specifically designed for Dutch language applications. It is built upon the robust Qwen1.5-7B-Chat base model and has undergone Direct Preference Optimization (DPO) using the Dutch BramVanroy/ultra_feedback_dutch_cleaned dataset.

Key Capabilities and Performance

• Dutch Language Proficiency: Optimized for natural language understanding and generation in Dutch.
• DPO Alignment: Enhanced through DPO finetuning for improved conversational quality and alignment with human preferences.
• ScandEval Performance: Achieves scores on the Dutch Natural Language Understanding and Dutch Natural Language Generation leaderboards that are very close to GPT-3.5's performance.
• Base Model Strengths: Inherits the general capabilities of the Qwen1.5-7B-Chat model, including a 32K context length.

Evaluation Highlights

While primarily focused on Dutch, the model also shows competitive English performance on the Open LLM Leaderboard, with an average score of 53.66. Notable scores include 76.03 on HellaSwag (10-Shot) and 62.38 on MMLU (5-Shot).

Intended Uses

This model is particularly well-suited for applications requiring high-quality Dutch language processing, such as chatbots, content generation, and language understanding tasks in Dutch. Users should be aware of potential biases and hallucinations inherent in LLMs and perform thorough validation for specific use cases.