ParaVT: Training Corpora for Long-Video Understanding Agents

ParaVT-Parquet provides two training corpora for a multi-agent framework designed for long-video understanding. The dataset includes a 97,000-row corpus for supervised fine-tuning and a 4,406-row corpus for reinforcement learning. The ParaVT project created this data, which was last updated on May 18, 2026.

Use Cases

• Training a video understanding agent based on the described supervised fine-tuning corpus.
• Applying reinforcement learning to video agents based on the described diverse RL corpus.
• Implementing the PARA-GRPO post-training method for multi-agent frameworks as described in the project overview.

Strengths

• Provides a substantial 97,000-row corpus specifically for cold-start supervised fine-tuning.
• Includes a separate, diverse reinforcement learning corpus of 4,406 rows.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is unknown, which may limit suitability assessment.

Provenance

Source

ParaVT project, hosted on Hugging Face.

Collection Method

Created as training annotations for the ParaVT multi-agent framework.

Freshness

Last updated 2026-05-18 11:50:23; freshness should be verified.