Spaces:
Sleeping
Apply for a GPU community grant: Academic project
I am building an open-source tool for verifiable code generation and execution using a fine-tuned Qwen-2.5-Coder-1.5B model. This project explores "verifiable execution rewards" by allowing users to pick a task, generate code, and immediately validate that code through automated tests in a secure sandbox.
I also wanted to show the capabilities of GRPO for verifiable rewards on bigger model Qwen 7B model for which i need GPU for inference.
Why this project needs a GPU grant:
Real-time Inference: To provide a responsive "watch generation + execution" experience, the model requires low-latency GPU inference (ZeroGPU or A10G).
Open Science/Source: The entire project, including training data and execution logic, is being released to the community to advance research in self-correcting code agents.
Community Education: This Space serves as a learning tool for developers to understand how reinforcement learning with execution feedback can improve 1.5B parameter models.
Current Progress:
The Gradio UI is set up, and the sandbox environment is ready. A GPU grant would allow me to make this demo public and stable for the community to explore.