Исследование структуры и качества датасетов

data

investigate-dataset is a Claude Code skill that explores the structure, fields, and data quality of datasets from HuggingFace, CSV, and JSON/JSONL files without relying on pre-written scripts. It guides the full investigation workflow: locating dataset configuration in eval code (variables like `DATASET_PATH`, `DATASET_REVISION`, and the `record_to_sample()` function), loading raw data via `datasets.load_dataset()`, `pd.read_csv()`, or `pd.read_json()`, and examining schema, shape, missing values, duplicates, and value distributions. Quick-reference `uv run python` commands allow inspecting HuggingFace dataset metadata without a full download, and the skill handles gated datasets via `HF_TOKEN`, rate limiting with retry backoff, and large datasets via streaming or split sampling. It is aimed at developers building or debugging Inspect AI evaluation pipelines who need to understand their data before writing transformation logic.