qihoo360/Light-R1-14B-DS

Light-R1-14B-DS: State-of-the-Art 14B Math Model

Light-R1-14B-DS, developed by Qihoo360, is a 14 billion parameter model derived from DeepSeek-R1-Distill-Qwen-14B. It represents a significant advancement as the first open-source model to successfully implement Reinforcement Learning (RL) on an already long-Chain-of-Thought (COT) fine-tuned model within a modest computational budget. This approach has led to notable improvements in mathematical reasoning capabilities.

Key Capabilities & Performance

• State-of-the-Art Math Performance: Achieves leading scores among 14B math models, with 74.0 on AIME24 and 60.2 on AIME25, surpassing many 32B models.
• RL Post-Training: Underwent a specialized long-COT RL Post-Training process, demonstrating expected behavior with simultaneous increases in response length and reward scores.
• Robustness: Performs well on the GPQA benchmark without specific training, indicating strong generalization.
• Data Decontamination: Features thorough data decontamination processes, including exact and N-gram matching, to ensure benchmark integrity.

Good For

• Advanced Mathematical Problem Solving: Ideal for applications requiring high-accuracy mathematical reasoning and complex problem-solving.
• Research in RL for LLMs: Provides a valuable open-source example of successful RL application on pre-fine-tuned models.
• Benchmarking and Development: A strong candidate for evaluating and developing new techniques in mathematical AI.