Datasets:
| name: ComputeEval | |
| description: >- | |
| ComputeEval is a benchmark for evaluating LLM-generated CUDA code on correctness | |
| and performance. It contains 591 handcrafted programming challenges — spanning kernels, | |
| runtime APIs, and GPU libraries — each with a held-out test harness for functional | |
| validation and optional performance benchmarks measuring GPU execution time against | |
| a baseline. Scored by Pass@1. | |
| evaluation_framework: nemo-evaluator | |
| tasks: | |
| - id: compute_eval | |
| config: default | |
| split: eval | |
| - id: compute_eval_2025_1 | |
| config: 2025-1 | |
| split: eval | |
| - id: compute_eval_2025_2 | |
| config: 2025-2 | |
| split: eval | |
| - id: compute_eval_2025_3 | |
| config: 2025-3 | |
| split: eval | |
| - id: compute_eval_2026_1 | |
| config: 2026-1 | |
| split: eval | |
| - id: compute_eval_2026_2 | |
| config: 2026-2 | |
| split: eval | |