mkurman/LiquidAI-LFM2.5-350M-SYNTH
mkurman/LiquidAI-LFM2.5-350M-SYNTH is a 350 million parameter reasoning model developed by mkurman, fine-tuned from LiquidAI/LFM2.5-350M. Utilizing a hybrid convolutional and sparse full-attention LFM2 architecture, it is optimized for mathematical, medical, and general reasoning tasks. This compact model produces explicit chain-of-thought reasoning and supports a 128k token context window, making it suitable for CPU or modest GPU deployment.
Loading preview...
Overview
mkurman/LiquidAI-LFM2.5-350M-SYNTH is a 350 million parameter language model developed by mkurman, specifically fine-tuned for reasoning tasks. It is based on the LiquidAI/LFM2.5-350M model and leverages a hybrid LFM2 architecture combining convolutional and sparse full-attention blocks. This design allows it to be compact and efficient, capable of running on CPUs or modest GPUs, while retaining a substantial 128,000 token context window.
Key Capabilities
- Reasoning-first: Trained on synthetic reasoning data, including math, medical, and general reasoning traces, to produce explicit chain-of-thought (
<think> ... </think>) before generating answers. - Compact & Fast: With approximately 350M parameters and a bf16 precision, it offers efficient performance.
- Long Context: Inherits a 128k
max_position_embeddingsfrom its base model. - Chat & Tools: Supports the ChatML format and includes native special tokens for tool-calling.
Training Details
The model was fully fine-tuned for 4,000 steps on a single A100 GPU, using a sequence length of 2,048 and an effective batch size of 64. Training utilized datasets such as PleIAs/SYNTH, mkurman/medical-reasoning-synthlabs-I, and mkurman/gsm8k-SynthLabs-reasoning.
Intended Use
This model is designed for research and experimentation with small reasoning models. Due to its size and synthetic training data, it may exhibit hallucinations and reasoning errors, and is not intended for high-stakes decisions.