Reynaerde-7B-Instruct: Dutch Chat Model

Reynaerde-7B-Instruct is a 7 billion parameter language model developed by Julien Van den Avenne, specifically fine-tuned for Dutch language chat. It is based on the powerful Mistral-7B-v0.3-Instruct architecture, leveraging its capabilities for conversational AI.

Key Capabilities and Features

• Dutch Language Specialization: The model has undergone extensive fine-tuning on several Dutch datasets, including ReBatch/ultrachat_400k_nl, BramVanroy/stackoverflow-chat-dutch, and BramVanroy/no_robots_dutch.
• Instruction-Tuned: Inherits instruction-following capabilities from its Mistral-7B-v0.3-Instruct base.
• Training Details: Fine-tuned using QLoRa in bfloat16 with Flash Attention 2 on an A100 PCIe, utilizing the sft script from the alignment handbook.
• Context Length: Supports a context window of 4096 tokens.

Intended Use Cases

This model is primarily intended for generating conversational responses and understanding prompts in the Dutch language. It is suitable for applications requiring a specialized Dutch chat model. Users should be aware of potential limitations, including the generation of incorrect, misleading, or offensive content, and should use the model with Mistral's chat template for optimal performance.