Multi-turn agentic traces of Claude Opus 4.8 converting PyTorch modules into
Triton GPU kernels. Each row is one problem from
GPUMODE/KernelBook: the model
writes a kernel, runs it on a GPU against the reference, reads the
correctness + speedup feedback, and iterates — so every trace is a grounded,
tool-using optimization loop, not a single-shot completion.
