{"data":{"kind":"file","path":"README.md","version_id":"o4gjh7gwt4gfd7wj6s5s0d7g","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":784,"modified_at":"2026-01-04T01:23:28.815000","content_hash":"84848db28a29da8c349b6de1f8bc1958d8f5b77300f0a95c7bc0cdd32d6d6d37"},"entries":[],"content":"# kernelbook-env\n\nEvaluates LLM-generated Triton GPU kernels against [KernelBook](https://huggingface.co/datasets/GPUMODE/KernelBook).\n\n## Task\n\nConvert PyTorch code to self-contained Triton kernel with `call(args)` entry point.\n\n## Required Format\n\n```python\nimport torch\nimport triton\nimport triton.language as tl\n\n@triton.jit\ndef triton_kernel(...):\n    ...\n\ndef call(args):\n    arg0_1, = args\n    ...\n    return (buf0,)\n\nclass ModuleNameNew(torch.nn.Module):\n    def forward(self, input_0):\n        return call([input_0])[0]\n```\n\n## Reward\n\n- `speedup = ground_truth_time / generated_time` if code passes\n- `0.0` if code fails (compile, runtime, or correctness error)\n\n## Usage\n\n```bash\nuv add verifiers datasets modal\nmodal setup\nvf-eval kernelbook-env -m gpt-4.1-mini -n 10\n```\n","encoding":"utf-8","truncated":false,"total_bytes":784},"status":null}