Veri-Code/ReForm-SFT-7B
Veri-Code/ReForm-SFT-7B is a 7.6 billion parameter language model developed by Veri-Code, specifically fine-tuned for formal software verification using Reinforcement Learning (RL) with LLMs. It focuses on generating and verifying code in formal languages like Dafny, aiming to reduce reliance on human-annotated priors. This model excels at producing syntactically valid and verifiable Dafny code, addressing scalability and reliability in program verification.
Loading preview...
ReForm-SFT-7B: Formal Software Verification with RL and LLMs
Veri-Code/ReForm-SFT-7B is a 7.6 billion parameter model from the Re:Form framework, designed for formal software verification. This model integrates Reinforcement Learning (RL) with Large Language Models (LLMs) to generate and verify code in formal languages, with a particular focus on Dafny. The core innovation lies in reducing the dependency on extensive human-annotated priors, a common challenge in informal language-based LLMs.
Key Capabilities
- Formal Verification: Utilizes RL strategies with feedback from a formal language verifier to ensure mathematically provable reasoning and outcomes.
- Dafny Code Generation: Specifically fine-tuned to generate syntactically valid and verifiable Dafny code.
- Scalable Data Curation: Employs an automatic and scalable data curation process for training.
- Benchmark Performance: Even smaller SFT models within the Re:Form framework have demonstrated superior performance in generating verifiable Dafny code compared to proprietary models on the DafnyComp benchmark.
- Reduced Human Priors: Aims to minimize the need for human-annotated data by grounding LLMs in rigorous formal systems.
Good For
- Automated Program Verification: Ideal for tasks requiring automatic and provable verification of software.
- Dafny Development: Generating and assisting with Dafny code creation.
- Research in Formal Methods: Exploring the application of LLMs and RL in formal software engineering and verification.
This model is compatible with the Hugging Face transformers library for text generation, particularly for Dafny code.