Datasets:
New analysis quickstart: explore 105,000 code generations without rerunning models
Tian Zhao and I have added a starter notebook and command-line example for Complexity Kink, the dataset accompanying our NeurIPS 2026 Evaluations and Datasets Track paper.
The main benchmark pairs 5,000 Python prompts with 105,000 generations across 21 evaluated models, prompt-side structural scores, test outcomes, and independent LLM audit results.
Start here: Analysis notebook | Python script
The examples show how to:
- Inspect the released model panel and benchmark coverage.
- Compare raw test outcomes with the audited and reviewed outcome definitions.
- Examine how construction-frame composition affects pooled complexity/performance patterns.
They use existing results, require no new model calls, and do not execute the released model-generated code. Results remain conditional on the benchmark and its historical settings; the dataset card explains the outcome definitions, intended uses, and limitations.
If you study prompt-side metrics, LLM judges, or benchmark validity, we would welcome independent analyses, reproductions, and feedback. Share your research question or an analysis link in this discussion.