Leanstral 1.5 Model Released with Enhanced Formal Verification Capabilities
Leanstral 1.5, a free model with 6 billion active parameters, has been launched, offering significant advancements in formal verification. It has solved numerous complex problems and identified previously unknown bugs in open-source code, demonstrating its practical applications in proof engineering.
Leanstral 1.5 is an open-source model licensed under Apache-2.0, featuring 6 billion active parameters and a total of 119 billion parameters.
The model has achieved state-of-the-art results, solving 587 out of 672 problems on the PutnamBench and reaching 87% on the FATE-H benchmark and 34% on the FATE-X benchmark.
It employs a three-stage training process that includes mid-training, supervised fine-tuning, and reinforcement learning with CISPO, allowing it to excel in proof engineering and real-world code verification.
Leanstral 1.5 has been evaluated on various benchmarks, including miniF2F, PutnamBench, FATE-H, and FATE-X, demonstrating superior performance compared to other models while being more cost-effective.
The model has successfully identified 5 previously unreported bugs across 57 tested repositories, showcasing its effectiveness in real-world applications.
Leanstral 1.5 is available for use via Hugging Face and as a free API, making it accessible for developers and researchers in the field of formal verification.