RLDriver/RLDriver-32B

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 8, 2026Architecture:Transformer Featherless Exclusive Cold

RLDriver/RLDriver-32B is a 32.8 billion parameter model, fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct using GRPO for fuzzing harness generation. It specializes in creating test harnesses for C/C++ libraries, trained across 10 diverse libraries like cJSON and curl. The model optimizes for coverage, alignment, throughput, and stateful API interaction, making it highly effective for automated software testing and vulnerability discovery.

Loading preview...

RLDriver/RLDriver-32B: GRPO-Fine-tuned for Fuzzing Harness Generation

RLDriver/RLDriver-32B is a specialized 32.8 billion parameter language model, derived from the Qwen/Qwen2.5-Coder-32B-Instruct architecture. Its core innovation lies in its fine-tuning using GRPO (Generalized Reinforcement Learning Policy Optimization), specifically for the task of generating fuzzing harnesses.

Key Capabilities

  • Automated Fuzzing Harness Generation: Designed to automatically create effective test harnesses for C/C++ libraries.
  • Targeted Training: Trained on a diverse dataset of 10 C/C++ libraries, including prominent ones like cJSON, curl, libjpeg, libtiff, libvpx, and zlib.
  • Multi-objective Optimization: Fine-tuned with a reward system that simultaneously optimizes for four critical metrics:
    • Coverage: Maximizing code paths explored.
    • Alignment: Ensuring generated harnesses are syntactically and semantically correct.
    • Throughput: Enhancing the efficiency of the fuzzing process.
    • Stateful API Interaction: Generating harnesses that correctly interact with complex, stateful APIs.

Good For

  • Software Security Testing: Ideal for developers and security researchers focused on automated vulnerability discovery in C/C++ applications.
  • Automated Test Case Generation: Useful for generating robust and effective test cases that go beyond typical unit tests.
  • Research in Reinforcement Learning for Code: Provides a practical application of GRPO in a code generation context.

This model offers a unique approach to improving software reliability and security by leveraging advanced reinforcement learning techniques to automate a complex and critical aspect of software testing. Further details on training scripts and the static-analysis backend are available in the code repository.

Finetunes
1 models