SN21 training bundle

Training bundle for the SN21 daily stream — 3069 settled episodes, one JSON object per line.

Each record is {episode_id, input, labels}: the input payload a miner sees at prediction time, and the settled outcomes at 7, 14 and 28 days (cost, conversions, CPA and conversion-value deltas).

What this is drawn from. This training set is drawn from the WEEKLY ERA and predates the 72-hour episode cap now live on the daily stream. Episode clustering therefore differs slightly from the baskets you will be scored on: it is the right task on a slightly different population — good for training, NOT a like-for-like preview of the daily stream. A held-out evaluation set from the same era is retained and never published.

Not scored. goal_miss_probability and instability_risk are accepted by the prediction contract for forward compatibility but are NOT SCORED and have no ground truth in this bundle: no outcome table carries them. Do not spend model capacity on them.

A held-out evaluation set is retained and never published, so a submitted model can be assessed on data it cannot have trained on.

How to use it: docs/SN21_TRAINING.md