NovachronoAI/LFM2.5-1.2B-Nova-Function-Calling
NovachronoAI/LFM2.5-1.2B-Nova-Function-Calling is a 1.2 billion parameter model based on Liquid AI's Liquid Neural Network (LFM 2.5) architecture, fine-tuned by NovachronoAI. This model is specifically engineered for robust function calling, enabling it to convert natural language queries into structured JSON inputs for tools and APIs. Despite its small size, it rivals larger 7B+ class models in function calling precision and efficiency, making it suitable for low-resource hardware.
Loading preview...
Model Overview
NovachronoAI/LFM2.5-1.2B-Nova-Function-Calling is a specialized 1.2 billion parameter model, fine-tuned from Liquid AI's LFM 2.5 Liquid Neural Network. This model is uniquely designed for robust function calling, efficiently translating natural language user requests into structured JSON for tool and API interactions. Its hybrid architecture allows it to achieve performance comparable to larger 7B+ class models on specific tasks, despite its significantly smaller footprint.
Key Capabilities
- Hyper-Efficient: Optimized for low-resource hardware like phones and Raspberry Pi due to its 1.2B Liquid architecture.
- Precision Tuned: Achieved a low training loss of 2.63, indicating mastery of structured JSON syntax without overfitting.
- Syntax Reliability: Demonstrates 97% syntax reliability in generating valid, crash-free JSON structures, matching GPT-4o class performance.
- ChatML Native: Utilizes the standard
<|im_start|>format for straightforward integration. - GGUF Ready: Available in various quantization levels, including an Imatrix GGUF version for high quality on low VRAM devices.
Training and Dataset
The model was trained on a curated subset of 15,000 high-complexity examples from the NovachronoAI/Nova-Synapse-Function-Calling dataset. This dataset, derived from over 130,000 complex user-agent interactions, emphasizes correct JSON schema adherence, argument extraction, and tool selection logic. The training process, using Unsloth and Hugging Face TRL, involved approximately 2 epochs, rapidly converging to a final loss of 2.63.