Anonymiserad/SweMedMCQ-LFM2.5-1.2B-Trace-SV
SweMedMCQ-LFM2.5-1.2B-Trace-SV is a 1.2 billion parameter language model developed by Anonymiserad, fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct with a 32768 token context length. It is specifically optimized for English and Swedish medical multiple-choice questions, utilizing reasoning-trace data in Swedish. This model excels at medical question answering, demonstrating improved accuracy over its base model, particularly in Swedish medical contexts, and is designed for research in medical LLM adaptation.
Loading preview...
Overview
SweMedMCQ-LFM2.5-1.2B-Trace-SV is a 1.2 billion parameter language model, fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct, specifically for medical multiple-choice questions in both English and Swedish. Its key differentiator is the use of reasoning-trace supervision during training, where it learns from detailed explanations and answers provided by larger teacher models (GPT-OSS-120B and Qwen3.6-35B-A3B). This approach aims to enhance the model's ability to generate reasoned responses.
Key Capabilities
- Medical Question Answering: Specialized in answering multiple-choice questions in a medical context.
- Bilingual Support: Proficient in both English and Swedish medical Q&A, with training data primarily in Swedish.
- Reasoning Trace Generation: Capable of producing a
<think>...</think>block for internal reasoning before providing a JSON-formatted answer, mimicking human thought processes. - Improved Accuracy: Achieves significantly higher accuracy on the SweMedMCQ test set (e.g., 49.8% in Swedish reasoning mode) compared to its base model (23.5% in Swedish).
Good for
- Research in Medical LLMs: Ideal for researchers exploring LLM adaptation for specialized domains and multilingual medical contexts.
- Benchmarking: Suitable for evaluating performance on medical multiple-choice question datasets, particularly the SweMedMCQ benchmark.
- Developing Medical Q&A Systems: Can serve as a foundation for applications requiring structured, explainable answers to medical queries, though it is a research model and not for clinical use.
Limitations
- This is a research model and is not intended for clinical use, diagnosis, or treatment decisions.
- Accuracy, while improved, is still below that of much larger models, and explanations can be incorrect.
- Swedish training and test data are machine-translated, which may introduce errors.