Behavior-1K 2025 Challenge V-JEPA 2 ViT-G Embeddings

V-JEPA 2 ViT-G Embeddings provide precomputed video features for a 62-hour subsample of the BEHAVIOR-1K 2025 challenge demonstrations. The dataset, created by quastAI, aims to accelerate downstream experimentation by eliminating repeated video processing. It was last updated on May 22, 2026.

Use Cases

• Train or fine-tune video understanding models based on precomputed V-JEPA 2 embeddings.
• Benchmark model performance on the BEHAVIOR-1K challenge tasks using standardized features.
• Conduct reproducible research in video-based AI by using a fixed set of extracted embeddings.
• Analyze video content and activity patterns based on the encoded visual-temporal features.

Strengths

• Contains embeddings for a 62-hour sample of video demonstrations.
• Uses the V-JEPA 2 ViT-G encoder, a specific and named model architecture.
• Explicitly aims to lower the barrier for teams by making experimentation faster and more reproducible.

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

huggingface

Collection Method

Precomputed from BEHAVIOR-1K 2025 challenge demonstrations using the V-JEPA 2 ViT-G encoder.

Freshness

Last updated 2026-05-22 15:32:05