project-numina/Kimina-Prover-RL-0.6B
AI-MO/Kimina-Prover-RL-0.6B is a 0.8 billion parameter theorem proving model developed by Project Numina and Kimi teams. Fine-tuned from AI-MO/Kimina-Prover-Distill-0.6B, it specializes in competition-style problem solving within the Lean 4 theorem prover environment. This model is trained using reinforcement learning and achieves 71.3% Pass@32 on MiniF2F-test, making it suitable for automated theorem proving tasks.
Loading preview...
AI-MO/Kimina-Prover-RL-0.6B: Reinforcement Learning for Lean 4 Theorem Proving
AI-MO/Kimina-Prover-RL-0.6B is a specialized 0.8 billion parameter model developed by Project Numina and Kimi teams, focusing on automated theorem proving in Lean 4. This model is an advancement over its base, AI-MO/Kimina-Prover-Distill-0.6B, having been fine-tuned through a reinforcement learning pipeline.
Key Capabilities & Performance
- Theorem Proving in Lean 4: Specifically designed and optimized for generating proofs within the Lean 4 formal verification system.
- Competition-Style Problem Solving: Excels at solving problems typically found in mathematical competitions, as demonstrated by its performance on MiniF2F-test.
- Reinforcement Learning (RL) Enhanced: The model's capabilities are boosted by reinforcement learning, leading to improved performance over its distilled predecessor.
- Benchmark Improvement: Achieves a 71.3% Pass@32 on the MiniF2F-test benchmark, representing a +2.45% improvement over the base model.
- Extended Context Length: Supports a substantial context length of 40960 tokens, beneficial for complex proofs.
Use Cases
- Automated Theorem Proving: Ideal for tasks requiring the automatic generation of formal proofs in Lean 4.
- Mathematical Research: Can assist researchers and mathematicians in formalizing and verifying mathematical statements.
- Educational Tool: Potentially useful for students and educators learning formal methods and theorem proving.
This model leverages an open-source training pipeline, Kimina-Prover-RL, and further details on its training methodology are available in this blog article.