Hugging Face has released Kimina-Prover-RL, an open-source training pipeline for formal theorem proving in the Lean 4 programming language. This pipeline, inspired by DeepSeek-R1 and utilizing the GRPO reinforcement learning approach with the Verl library, enables models to generate natural language reasoning traces followed by corresponding Lean code. Two models have been released as a result: AI-MO/Kimina-Prover-RL-1.7B and AI-MO/Kimina-Prover-RL-0.6B, both setting new state-of-the-art benchmarks on the MiniF2F dataset for their respective size categories. AI
IMPACT This release provides a new open-source tool for training AI models in formal theorem proving, potentially advancing capabilities in mathematical reasoning and code generation.
RANK_REASON Release of an open-source training pipeline and two models for formal theorem proving, setting new benchmarks.
- Claude
- AI-MO/Kimina-Prover-RL-0.6B
- AI-MO/Kimina-Prover-RL-1.7B
- DeepSeek-R1
- kimina-client
- kimina-lean-server
- Kimina-Prover-RL
- Lean 4 Programming Language
- miniF2F
- NuminaMath-LEAN
- Verl
- Claude 2.0
- GRPO
AI-generated summary · Google Gemini · from 6 sources. How we write summaries →
