User01110/LFM-2.5-350M-MathMini
User01110/LFM-2.5-350M-MathMini is a 350 million parameter language model, fine-tuned from LiquidAI/LFM2.5-350M, specifically optimized for mathematical problem-solving. Trained on a curated math dataset with a 32768 token context length, this model is designed to act as a math-focused assistant. It specializes in generating structured responses to math queries, following a specific format for understanding, drafting, refining, and providing a final answer.
Loading preview...
User01110/LFM-2.5-350M-MathMini: Math-Focused Language Model
This model is a 350 million parameter language model, fine-tuned from LiquidAI/LFM2.5-350M, with a focus on mathematical problem-solving. It was trained using the TRL SFTTrainer on the User01110/math-curated-dataset, comprising 39,040 training rows over one epoch.
Key Capabilities and Training Details
- Specialized Math Assistant: The model is designed to function as a math-focused assistant, guided by a system prompt that emphasizes a structured problem-solving approach: "Understanding Query, Drafting Answer, Refining The Answer, and Final Response."
- Training Format: It was trained to generate full responses based on a chat template, with a maximum sequence length of 1024 tokens.
- Context Length: The base model supports a 32768 token context length.
Important Limitation
It is crucial to note that this model was trained on generated math-style data. Consequently, its responses may contain incorrect arithmetic or flawed reasoning. Users should independently verify any mathematical answers provided by the model, as it should not be treated as a reliable source for accurate mathematical solutions without verification.
Usage
The model can be easily integrated using the transformers library for generating responses to math problems, following the specified chat template and system prompt.