ParaVT-Source: Raw Video Archives for Multi-Agent Long-Video Understanding

ParaVT provides source media archives for the ParaVT training corpus. The dataset bundles raw video files referenced by the ParaVT-Parquet annotations, packaged as per-source zip archives. It was created by ParaVT and last updated on May 18, 2026.

Use Cases

• Training multi-agent video understanding models based on the raw video files.
• Post-training models with PARA-GRPO based on the referenced video archives.
• Benchmarking long-video understanding systems based on the paired source and annotation data.

Strengths

• Files are grouped by sentinel bucket, which may aid organization.
• Dataset is explicitly paired with a separate, detailed annotation repository (ParaVT-Parquet).

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

ParaVT

Collection Method

Source media archives bundled from referenced raw video files.

Freshness

Last updated 2026-05-18 11:50:58; freshness should be verified.