Anonymiserad/SweMedMCQ-LFM2.5-1.2B-Answer-EN
Anonymiserad/SweMedMCQ-LFM2.5-1.2B-Answer-EN is a 1.2 billion parameter language model, fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct, specifically for answering English medical multiple-choice questions. It was trained on 135,093 English medical questions with answer-level supervision, targeting the gold option alone. This model is designed for medical question answering tasks, particularly in an exam-style multiple-choice format, and is part of a research effort to benchmark and adapt LLMs for English-Swedish medical QA.
Loading preview...
SweMedMCQ-LFM2.5-1.2B-Answer-EN Overview
This model is a 1.2 billion parameter language model, fine-tuned from the LiquidAI/LFM2.5-1.2B-Instruct base model. It is specifically designed for answer-level supervision on English medical multiple-choice questions, meaning it predicts only the correct option (e.g., "B") rather than generating full explanations. This model is one of 14 student models developed for the "Benchmarking and Adapting LLMs for English–Swedish Medical Question Answering" paper.
Key Capabilities
- Specialized Medical QA: Fine-tuned exclusively on 135,093 English medical multiple-choice questions from the MedicalAgentsBench datasets.
- Answer-Level Prediction: Focuses on directly identifying the correct answer option in a structured JSON format (
{"answer_index": "<letter>"}). - Research-Oriented: Developed as part of a research initiative to evaluate LLMs in medical question-answering contexts.
Good For
- Medical Exam Simulation: Ideal for tasks requiring direct, multiple-choice answer selection in English medical contexts.
- Benchmarking Medical LLMs: Useful for researchers evaluating model performance on specific medical QA datasets.
- Structured Output Generation: Excels at producing structured JSON outputs for answer indices.
Limitations
It is important to note that this is a research model and not a medical device; it should not be used for clinical decisions. Its accuracy is lower than larger models, and explanations, even if provided by a different model, might be incorrect. It is trained and evaluated solely on exam-style multiple-choice questions.