AtomicLM Skill Annotations: Real-World Cooking Video Steps

Self-induced, canonicalized atomic-skill annotations for real-world instructional cooking videos from wikiHow and YouTube. The annotations were curated by AutoMark and include verb(object) calls with second-aligned intervals over a 48-skill canonical library. This dataset is a companion to robot-side annotations and was last updated on 2026-06-03.

Use Cases

• Training robotic skill models based on verb(object) annotations aligned to video intervals.
• Analyzing the structure of instructional tasks from real-world cooking videos.
• Developing multimodal models that link visual steps in videos to canonical skill descriptions.
• Benchmarking skill recognition and segmentation algorithms in the cooking domain.

Strengths

• Annotations are based on a canonical and audited library of 48 skills.
• Annotations include second-aligned temporal intervals for precise step localization.
• Companion dataset exists for robot-side skill annotations, enabling cross-modal work.

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, author Chenwei1999

Collection Method

Self-induced, canonicalized annotations curated by AutoMark from wikiHow and YouTube instructional cooking videos.

Freshness

Last updated 2026-06-03 18:34:16; freshness should be verified.