Rumiii/LiquiMedThink1.2B
Rumiii/LiquiMedThink1.2B is a 1.17 billion parameter language model fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking, specifically adapted for medical chain-of-thought reasoning. It retains the base model's ability to produce explicit, step-by-step thinking traces before delivering a final answer. This model excels at clinical reasoning demonstrations and research into small language model capabilities in healthcare, utilizing a 4096 token context length.
Loading preview...
LiquiMedThink 1.2B: Medical Reasoning with Explicit Thinking
LiquiMedThink 1.2B is a 1.17 billion parameter model, fine-tuned from LiquidAI/LFM2.5-1.2B-Thinking, to specialize in medical chain-of-thought reasoning. A key differentiator is its ability to generate explicit, step-by-step reasoning traces (<think>...</think>) before providing a concise final answer. This behavior was successfully preserved through supervised fine-tuning on the FreedomIntelligence/medical-o1-reasoning-SFT dataset, which features genuinely separate reasoning traces and final answers, unlike prior attempts that lost or mimicked thinking ability.
Key Capabilities & Features
- Medical Chain-of-Thought Reasoning: Generates detailed, exploratory clinical thinking processes.
- Explicit Reasoning Traces: Outputs visible step-by-step thought processes, including uncertainty and self-correction.
- Domain Adaptation: Acquired clinical domain knowledge through fine-tuning on 19,704 medical reasoning samples.
- 4096 Token Context: Supports longer clinical questions and reasoning sequences.
- Efficient Fine-tuning: Utilized QLoRA (4-bit) via Unsloth, with only 0.78% trainable parameters.
Intended Use Cases
- Research: Investigating small language model capabilities in healthcare and clinical reasoning.
- Medical Education: Demonstrating clinical reasoning processes.
- Prototyping: Evaluating medical AI pipelines.
- Data Format Research: Studying the impact of fine-tuning data format on reasoning behavior.
Important Note: This model is not intended for clinical decision support or any high-stakes medical applications without expert oversight due to potential clinical inaccuracies and its exploratory thinking traces.