jump-gemma-adapter / DATASET_CARD.md
JUMP Artifact Maintainer
Publish verified sanitized model artifacts
51adac7
|
Raw
History Blame Contribute Delete
562 Bytes

Dataset and generator record

The pilot dataset was generated by the deterministic six-object Track H simulator at seed 7613. It contains 200 training and 24 held-out episodes. Raw dataset SHA-256 is 10b3d9cad161d415914e62795136675339a29c3cb141a4169b413da3dad75802; derived tokenized dataset SHA-256 is 1bfe1a46d19c8397e58f9fb8b4553524a0dce3764659966d1a1ec3c63b5d83ae. The 640-token policy reserves all 33 target-plus-EOS tokens and truncates only prompt tokens. The audit records every example and condition. This synthetic dataset contains no user data.