Matter-0.1-Slim-7B-preview: A Function-Calling Optimized Mistral Finetune
The 0-hero/Matter-0.1-Slim-7B-preview is a 7 billion parameter model, built upon the Mistral 7B architecture through a comprehensive full-finetuning process. Its training leveraged the slim-D version of the Matter dataset, which aggregates and analyzes over 6 billion tokens from more than 35 distinct datasets.
Key Capabilities
- Advanced Function Calling: This model is specifically engineered to support function calling, incorporating dedicated tokens (
<|begin_func|>, <|end_func|>, <|begin_func_response|>, <|end_func_response|>) for seamless integration with external tools and APIs. This enables it to generate structured function calls and process their responses effectively. - ChatML Prompt Format: It adheres to the ChatML prompt format, ensuring compatibility and ease of use within chat-based applications and conversational AI systems.
- Efficient Training: The model underwent 3 epochs of full-finetuning in approximately 17 hours using Axolotl on 4x A100 GPUs (80GB).
Ideal Use Cases
- Tool-Augmented AI: Excellent for applications where the LLM needs to interact with external systems, databases, or APIs via function calls.
- Intelligent Agents: Suitable for building agents that can perform actions, retrieve real-time information, or automate tasks by invoking specific functions.
- Structured Output Generation: Its function calling capabilities make it adept at generating structured outputs that can be directly parsed and utilized by other software components.