Document dataset scope, splits, EEF52, three-camera reader usage, project links, the owner-selected MIT license and the complete citation. Rename the reader to DexOmniDataset and refresh file checksums. Recording bytes, numeric values, split membership, deterministic language sampling and preview order remain unchanged.
Prioritize custom DexOmni tasks in episode previews
Place ten distinct custom implementations first. Swap two 72-frame training episodes to keep the default Episode 0 preview representative, with matching numeric identifiers, per-episode statistics, source manifests and hashes. All recordings, action/state values, source identities, split memberships, frame offsets and normalization are preserved. Historical render provenance remains unchanged and is linked through the source manifest. See provenance/HF_PREVIEW_ORDER.json.
