KKcoding2025/evo-eval_data

KramaBench (evo-eval krama_full)

End-to-end data-science agent benchmark, packaged for the evo-eval evaluation framework (evo_eval.dataset.v1 generic dataset schema).

Source: KramaBench (Lai et al., 2025), https://github.com/mitdbg/kramabench. If you use this data, please cite the upstream preprint:

@misc{lai2025KramaBench,
  title  = {KramaBench: Evaluating End-to-End Data-Science Agents},
  author = {Eugenie Lai and Gerardo Vitagliano and Ziyu Zhang and *et al.*},
  year   = {2025},
}

Layout (repo root = evo-eval DATA_ROOT)

• krama_full/tasks.jsonl — 104 top-level tasks, integer ids 1..104 (one per line: task_id, query.text, assets_dir, meta)
• krama_full/references/{task_id}.json — golden answers with answer_type (numeric_exact / string_exact / list_exact / numeric_approximate / list_approximate / string_approximate)
• krama_full/assets/{domain}/ — per-domain input data (copied from upstream data/{domain}/input)
• krama_full/subtasks/ — 631 per-step subtasks (N-K ids derived from parent task N), for intermediate-step diagnostics; not part of the official top-level metric

Notes

• Original KramaBench task ids (e.g. legal-hard-1) are preserved in meta.source_task_id (tasks) and meta.source_subtask_id (subtasks).
• Deterministic mapping: top-level tasks sorted by (domain, source_task_id), then numbered 1..104; subtask K of task N gets id N-K.
• Scoring (official metric dispatch by answer_type → success / f1 / f1_approximate / rae_score / llm_paraphrase) ships with the evo-eval dataset template configs/datasets/krama_full/, not with this data repo.