Anonymiserad/SweMedMCQ-LFM2.5-1.2B-Trace-EN

SweMedMCQ-LFM2.5-1.2B-Trace-EN Overview

This model, Anonymiserad/SweMedMCQ-LFM2.5-1.2B-Trace-EN, is a 1.2 billion parameter language model fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct. It is specifically designed for medical multiple-choice question answering in both English and Swedish, leveraging a unique reasoning-trace training methodology. This approach involves supervising the model with detailed reasoning steps from teacher models (GPT-OSS-120B and Qwen3.6-35B-A3B) before the final answer, enhancing its ability to provide structured explanations.

Key Capabilities

• Medical Question Answering: Specialized in answering exam-style medical multiple-choice questions.
• Reasoning-Trace Generation: Produces a <think>…</think> reasoning block before outputting a JSON-formatted answer, detailing its thought process.
• Multilingual Support: Fine-tuned for both English and Swedish medical contexts, though training data for reasoning traces was exclusively English.
• Structured Output: Generates answers in a consistent JSON format including an explanation and the answer index.

Good For

• Medical AI Research: Ideal for researchers exploring LLM performance in medical domains, particularly for diagnostic reasoning and explanation generation.
• Benchmarking: Part of a larger release for the SweMedMCQ benchmark, useful for evaluating and comparing LLMs on medical question-answering tasks.
• Educational Tools: Can be used to develop tools that help medical students practice and understand complex medical concepts through explained answers.

Limitations

It is crucial to note that this is a research model and not a medical device. It should not be used for clinical decisions. Its accuracy, while improved over the base model, is still below that of larger models, and explanations can sometimes be incorrect even if the final answer is right. The Swedish training and test data are machine-translated, which may introduce errors.