ermiaazarkhalili/LFM2.5-350M-Function-Calling-xLAM-Unsloth

Overview of LFM2.5-350M-Function-Calling-xLAM-Unsloth

This model, developed by ermiaazarkhalili, is a specialized fine-tune of the LiquidAI LFM2.5-350M base model, specifically optimized for function calling. It leverages the Unsloth framework, which enabled 2x faster training and 60% less VRAM usage during its development.

Key Capabilities & Features

• Function Calling Expertise: Fine-tuned on the extensive Salesforce/xlam-function-calling-60k dataset, comprising 60,000 examples of queries, tool definitions, and structured answers.
• Efficient Training: Utilizes Unsloth and QLoRA (4-bit) for efficient SFT, achieving a final training loss of 0.6507 with a peak VRAM of only 5.73 GB on an NVIDIA H100 GPU.
• Compact Size: At 350 million parameters, it offers a lightweight solution for integrating function calling capabilities into applications.
• Flexible Deployment: Available in standard Transformers format, and also provides GGUF versions for CPU and edge inference, compatible with tools like Ollama and llama.cpp.
• English Language Focus: Primarily trained on English data, making it best suited for English-language function calling tasks.

Ideal Use Cases

• Tool Use & Agent Systems: Excellent for applications requiring the model to interpret user requests and generate appropriate function calls to external tools or APIs.
• Resource-Constrained Environments: Its small size and efficient training/inference (especially with Unsloth and GGUF) make it suitable for deployment where computational resources are limited.
• Prototyping & Development: A strong candidate for developers looking to quickly integrate function calling without the overhead of larger models.