LiquidAI/LFM2-350M-Math
LiquidAI's LFM2-350M-Math is a 350 million parameter reasoning model, based on LFM2-350M, specifically designed to excel at solving complex mathematical problems. This compact model leverages advanced reasoning techniques to structure thought processes and self-verify solutions, making it highly capable for its size in tackling challenging math tasks. It is optimized for edge deployment, focusing on limiting memory consumption and latency while maintaining high performance in mathematical reasoning.
Loading preview...
LFM2-350M-Math: A Compact Reasoning Model for Mathematics
LFM2-350M-Math, developed by LiquidAI, is a 350 million parameter model built upon the LFM2-350M architecture. It is specifically engineered as a "tiny reasoning model" to address intricate mathematical problems effectively. The model's design emphasizes structured thought processes, exploration of multiple solution strategies, and self-verification of responses, enabling it to solve challenging competition-level math problems despite its small size.
Key Capabilities & Features
- Specialized Mathematical Reasoning: Highly capable for its size in solving complex math problems.
- Optimized for Edge Deployment: Designed to limit memory consumption and latency, making it suitable for resource-constrained environments.
- Concise Reasoning: Leverages a post-training recipe, including reinforcement learning with explicit reasoning budgets and difficulty-aware advantage re-weighting, to reduce response verbosity without sacrificing accuracy.
- Single-Turn Conversations: Intended for single-turn interactions.
- English Language Support: Currently supports English only.
Performance & Usage
Benchmark evaluations demonstrate LFM2-350M-Math's strong performance in response accuracy for mathematical tasks. For optimal generation, LiquidAI recommends using greedy decoding with specific parameters (temperature=0.6, top_p=0.95, min_p=0.1, repetition_penalty=1.05) and no system prompt. The model utilizes a ChatML-like chat template for interaction. Colab notebooks are provided for easy inference and fine-tuning using methods like SFT (TRL, Axolotl, Unsloth) and DPO (TRL).