mkurman/lfm2-350M-med
The mkurman/lfm2-350M-med is a 0.35 billion parameter language model, fine-tuned by mkurman on top of LiquidAI's LFM2-350M base. This model specializes in medical question-answering and tool-augmented search, designed for efficient inference on laptops and edge devices. It incorporates supervised fine-tuning on open-source medical data and tool-calling traces, followed by DPO preference alignment using MedMCQA, making it suitable for medical research and educational applications.
Loading preview...
Model Overview
mkurman/lfm2-350M-med is a specialized 350 million parameter language model, fine-tuned from LiquidAI's LFM2-350M base. It is designed for medical Q&A and tool-augmented search, optimized for on-device inference with strong CPU latency. The model supports tool use with dedicated special tokens, following a ChatML-like template.
Key Specialization Steps
- Domain SFT: Instruction-style Q&A from open medical sources and synthetic conversions, alongside tool-use (search) supervised traces.
- Preference Alignment: Direct Preference Optimization (DPO) using MedMCQA-derived preferences to favor clinically reasonable short answers.
- Model Fusion: Utilizes Arcee Fusion in MergeKit for controlled weight fusion, avoiding over-averaging.
Performance & Limitations
Evaluated on an author's harness, the model achieves 19.46 on MMLU-Pro (vs. 18.76 for the base) and 52.595 on IFEVAL (vs. 61.72 for the base). It is intended for education and research purposes only. Users are cautioned that this model is not a clinician; it may hallucinate and should not be used for diagnosis or treatment, emphasizing the need for qualified medical supervision.