mkurman/llama-3.2-MEDIT-3B-o1

Model Overview

mkurman/llama-3.2-MEDIT-3B-o1 is a 3 billion parameter Small Language Model (SLM) fine-tuned by mkurman from MedIT Solutions Llama 3.2 3B Instruct. Its core differentiation lies in its o1-like reasoning capabilities, achieved through the use of explicit <Thought> and <Output> tags. This structure encourages a chain-of-thought style of text generation, making it particularly suitable for tasks requiring structured reasoning.

Key Characteristics & Usage

• Reasoning Focus: Designed for instruct-style reasoning tasks, leveraging <Thought> and <Output> tags to separate internal reasoning from the final answer.
• Deterministic Outputs: Recommended for use with do_sample=False or temperature=0.0 to achieve more deterministic and exact matching outputs rather than diverse generation.
• Base Model: Built upon the Llama 3.2 3B Instruct architecture, licensed under llama3.2.
• Intended Use Cases:
  • General question answering
  • Instruction-based generation
  • Exploration of reasoning and chain-of-thought processes

Important Usage Notes

• Stop Sequences: Users must configure </Output> as a stop sequence during generation to prevent infinite output.
• Prompt Workaround: A known issue where the model might start with <|python_tag|> can be mitigated by appending "<Thought>\n\n" to the end of the generation prompt.

Limitations

• Hallucination: Like many LLMs, it may generate plausible but incorrect information.
• Medical Information: Not to be relied upon for sensitive, real-world medical diagnosis or advice without expert verification. It is not a certified medical professional.
• Bias: Outputs may reflect biases present in its training data.