Datasets:
ArXiv:
License:
Evaluator-only combined episodes
These JSONL files combine each episode's inputs and target values in one record for the scorer and offline statistical baselines:
MODEL_DEV67_EPISODES_V1_WITH_TARGETS.jsonlMODEL_TEST211_EPISODES_V1_WITH_TARGETS.jsonl
Use inputs/ for agent runs. Keep targets/ and these combined records in the
evaluation environment: they contain ground-truth values such as
future_counts and future_share that must not enter model prompts.