Safetensors
English
qwen2

LeanFormalizer_SFT

Based on Qwen2.5-7b and trained on our LeanStatement_SFT dataset, our model achieves state-of-the-art performance in formal mathematics verification, with a pass@1 compilation success rate of 94.1% (459/488) on MiniF2F and 75.4% (282/374) on ProofNet benchmarks as of December 2024(while our PPO model achieves 95.3% on MiniF2F and 76.0% on Proofnet). These results demonstrate the effectiveness of scaling laws in mathematical formalization using the Lean theorem prover.

Downloads last month
30
Safetensors
Model size
8B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SJTULean/LeanFormalizer_SFT

Base model

Qwen/Qwen2.5-7B
Finetuned
(3,273)
this model
Finetunes
1 model

Dataset used to train SJTULean/LeanFormalizer_SFT