Xperience-10M 128-Episode Metadata-Balanced Selection

This is a download plan, not a trained model result. It uses Hugging Face file metadata only and downloads no raw episode data.

Why This Selection

Use only complete episodes: annotation.hdf5 plus six MP4 streams.
Exclude visualization.rrd from the training download plan.
Avoid tiny annotation outliers that are likely one-segment examples.
Use one episode per top-level session to reduce leakage and overfitting to one capture session.
Balance across four annotation-size bands as a proxy for duration/content richness before category labels are available.
Split by session into train/val/test.

Selection Summary

Measure	Value
Selected episodes	128
Unique sessions	128
Split counts	{"test": 16, "train": 96, "val": 16}
Size-band counts	{"long": 32, "lower_mid": 32, "short": 32, "upper_mid": 32}
Estimated training download, no RRD	277.71 GiB
Estimated annotation bytes	226.53 GiB
Estimated windows at 256/episode	32768
Session leakage train/val	0
Session leakage train/test	0
Session leakage val/test	0

Filters

Rule	Value
Available complete episodes	12102
Candidates after filters	11478
Minimum annotation size	992.76 MiB
Minimum training size	1.22 GiB
Rejected counts	{"annotation_too_small": 606, "training_too_small": 18}

Split x Size Band

Split	short	lower_mid	upper_mid	long
train	24	24	24	24
val	4	4	4	4
test	4	4	4	4

Important Limitation

HF metadata does not expose semantic content categories. This selection is the best first-pass balance before downloading. After the selected annotations are staged, parse Main Task, Sub Task, Current Action, objects, and interaction text; then swap episodes if one content cluster dominates.

Output Files

JSON: results/omni_finetune/xperience10m_128_episode_selection.json
CSV: results/omni_finetune/xperience10m_128_episode_selection.csv
Download file list: results/omni_finetune/xperience10m_128_episode_download_files.txt