Behavior2025 Vjepa2 Vitg 130H Demo Embeddings: Precomputed Video Features

130 hours of video demonstrations from the BEHAVIOR-1K 2025 challenge, processed into feature representations. The embeddings were extracted using the V-JEPA 2 ViT-G encoder from Facebook Research. The dataset was created by quastAI and last updated on Hugging Face in May 2026.

Use Cases

• Benchmarking video representation models based on the BEHAVIOR-1K challenge demonstrations.
• Training downstream classifiers for activity recognition using the precomputed V-JEPA 2 features.
• Analyzing the semantic structure of complex, goal-oriented human behaviors in video.
• Developing and evaluating few-shot or transfer learning methods for video tasks.

Strengths

• Features are precomputed using the V-JEPA 2 ViT-G encoder, a state-of-the-art video representation model.
• The source material is a 130-hour subsample of demonstrations from the structured BEHAVIOR-1K 2025 challenge.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset contains only precomputed features, not the original video or audio data.

Provenance

Source

Derived from behavior-1k/2025-challenge-demos on Hugging Face.

Collection Method

Feature extraction performed using the V-JEPA 2 ViT-G encoder from the facebookresearch/vjepa2 repository.

Freshness

Last updated 2026-05-21 12:04:27