PIE2F-LongHorizon

A large-scale dataset of 147 long-horizon screen recording sessions (~646 hours) covering real computer workflows across IDEs, terminals, browsers, research, notes, and experiment monitoring.

PIE2F-LongHorizon preserves complete long-form sessions and provides structured temporal indexes, sampled visual activity labels, quality signals, and session metadata for studying extended computer use.

Dataset Summary

Sessions	Hours	Temporal Segments
147	646.21	11,696

Data

Each session contains a long-form screen recording together with structured metadata and temporal indexes.

The release includes:

• 147 complete session videos
• 646.21 hours of computer workflows
• 11,696 temporal activity segments
• timeline samples every 30 seconds
• weak visual activity labels
• quality measurements
• session metadata
• content hashes

Data Organization

data/videos/
indexes/sessions.parquet
indexes/segments.parquet
indexes/timeline_samples.parquet
indexes/review_queue.parquet
indexes/activity_taxonomy.json
indexes/data_contract.json
metadata/videos/

Intended Use

PIE2F-LongHorizon is designed for research on:

• long-horizon computer-use models
• workflow understanding
• temporal representation learning
• long-context retrieval
• video understanding
• workflow segmentation
• visual state retrieval
• further annotation of computer workflows

Annotations

The session timelines are sampled every 30 seconds and assigned weak visual activity labels.

Nearby predictions are temporally smoothed and grouped into activity segments, producing a lightweight index over the complete 646-hour dataset.

These labels are intended for navigation, retrieval, sampling, and further annotation rather than as human-verified task descriptions.

Notes

PIE2F-LongHorizon contains visual workflow data and does not include synchronized mouse or keyboard actions.

Activity labels are automatically generated weak annotations and may contain errors.

License

CC BY 4.0