Fate-Zero/Archer2.0-Code-1.5B-Preview
Fate-Zero/Archer2.0-Code-1.5B-Preview is a 1.5 billion parameter language model developed by Fate-Zero, featuring a 32K context length. It introduces Asymmetric Importance Sampling Policy Optimization (ASPO) to enhance reinforcement learning capabilities, specifically mitigating issues like entropy collapse and repetitive outputs. This model is particularly optimized for code generation and reasoning tasks, demonstrating improved performance on LiveCodeBench benchmarks compared to its predecessor and other 1.5B models.
Loading preview...
Archer2.0-Code-1.5B-Preview: Enhanced Reasoning for Code Generation
Archer2.0 represents a significant advancement in reinforcement learning for large language models, developed by Fate-Zero. This 1.5 billion parameter model, with a 32K context length, introduces Asymmetric Importance Sampling Policy Optimization (ASPO). ASPO is designed to overcome limitations of traditional PPO-Clip methods, effectively mitigating issues such as entropy collapse, repetitive outputs, and premature convergence, thereby enabling more sophisticated reinforcement learning capabilities.
Key Capabilities & Differentiators
- Advanced Reinforcement Learning: Utilizes ASPO to improve model stability and output quality, addressing common pitfalls in RL-trained LLMs.
- Optimized for Code Generation: Specifically evaluated and improved for coding tasks, as demonstrated by its performance on LiveCodeBench.
- Competitive Performance: Outperforms its predecessor, Archer-Code-1.5B, and other 1.5B parameter models like DeepSeek-R1-1.5B, DAPO, DeepCoder-1.5B, and Nemotron-1.5B on LiveCodeBench v5 and v6 benchmarks.
Ideal Use Cases
- Code Generation: Excels in generating functional and diverse code snippets.
- Code Reasoning Tasks: Suitable for applications requiring logical problem-solving within a coding context.
- Research in RL for LLMs: Provides a strong baseline and a novel approach (ASPO) for researchers exploring reinforcement learning techniques in language models.