Source-of-truth eval/train split for the data-agent RL pipeline.
Derived from jupyter-agent/jupyter-agent-dataset by:
Filtering to executor_type == "e2b" (29,555 rows survive; executor_type == "llm" rows have mismatched Kaggle metadata and are dropped).
Per-row classification of the gold answer into a default reward grading mode (see reward_mode_initial).
Stratified sampling by (reward_mode_initial × package_tier) with a max-K-per-Kaggle cap on the eval… See the full description on the dataset page:
https://huggingface.co/datasets/AdithyaSK/data_agent_rl.
