glm-5.2-kernelgym-rollouts

# GLM-5.2 KernelGym Rollouts This dataset contains 3,200 feedback-driven GPU-kernel optimization trajectories generated by `zai-org/GLM-5.2-FP8`: 100 validation tasks, two backends (inline CUDA and Triton), and 16 rollouts per task. Each trajectory retains the prompt/feedback message history, model responses and reasoning, extracted kernel code, KernelGym compilation and correctness results, profiling metadata, token usage, and stopping decision. Every published record ended with the model-controlled final decision and passed correctness, custom kernel execution, and nonnegative runtime validation. ## Collection configuration - Source tasks: `hkust-nlp/drkernel-validation-data`, validation split - KernelGym: https://github.com/hkust-nlp/KernelGYM at `3a84417f8c0efaadb215ef638b37d12e71ed20f3` - Model: `zai-org/GLM-5.2-FP8` - Sampling: temperature 1.0, top-p 0.95, up to 16,384 generated tokens - Context length: 262,144 tokens - Maximum feedback turns: 128 ## Quality summary | Backend | Records | Tasks | Mean turns | Median speedup | Faster than baseline | Unique final code | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | CUDA | 1,600 | 100 | 1.889 | 1.072x | 75.4% | 100.0% | | Triton | 1,600 | 100 | 1.976 | 1.427x | 88.1% | 100.0% | Speedups below 1.0 are retained because the dataset records valid optimization trajectories, not only improvements. ## Schema The compressed JSONL shards preserve the original nested rollout records. Core fields include `dataset_index`, `problem_id`, `problem_name`, `reference_code`, `backend`, `rollout`, `messages`, `turns`, `stop_reason`, and `final_response`. Each turn contains the model response and reasoning, extracted kernel code, evaluation results, usage, and elapsed time. ## Provenance and licensing The records include source task text and reference implementations from `hkust-nlp/drkernel-validation-data`. Review that dataset, KernelGym, and GLM-5.2 terms before redistribution or commercial use.