Per-version paired human + LLM-agent trajectories on 134 Kaggle competitions. Each row in state.parquet is one version of one ML coding session; each row in action.parquet is one transition between adjacent versions, with multi-label state / action / intent annotations and held-out Kaggle leaderboard scores.

Quickstart

from datasets import load_dataset

Paired split: 7 competitions with both human kernels and LLM-agent runs (head-to-head)

state_paired =… See the full description on the dataset page: https://huggingface.co/datasets/TraceML-HF/TraceML.