Qualcomm Interactive Cooking Dataset: Counterfactual Mistake Annotations for 1,110 Videos

Qualcomm's synthetic dataset contains 25,087 mistake-intervention annotations for interactive cooking guidance. It includes video segments paired with instruction and feedback text and their timestamps. The data is derived from the CaptainCook4D, Ego4D, and Ego-Exo4D source datasets.

Use Cases

• Training models for mistake detection in instructional videos based on annotated video-text pairs.
• Developing counterfactual reasoning systems for interactive guidance using the provided intervention annotations.
• Benchmarking video-language models on the task of generating corrective feedback for procedural errors.

Strengths

• Contains 25,087 annotated rows, providing a substantial number of examples.
• Derived from 1,110 unique videos across three established source datasets (CaptainCook4D, Ego4D, Ego-Exo4D).
• Includes synchronized timestamps for video segments and text pairs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Qualcomm, derived from CaptainCook4D, Ego4D, and Ego-Exo4D datasets.

Collection Method

Synthetic annotations created for video segments.

Freshness

Last updated 2026-05-05 14:57:50; freshness should be verified.