EgoCoT-Bench provides 3,172 samples across 351 unique videos for evaluating reasoning in first-person video. The benchmark, authored by DStardust and released in April 2026, is structured for evaluation with 300 public development samples and 2,872 public test samples. Its focus is on grounded and verifiable reasoning tasks within egocentric video contexts.
Use Cases
- Benchmarking model performance on grounded reasoning tasks based on the described evaluation structure
- Developing models for verifiable reasoning in first-person video based on the benchmark's stated focus
- Training or fine-tuning models for egocentric video understanding based on the 351 unique videos
- Analyzing multimodal reasoning capabilities based on the combination of video and likely textual or question-answer data
Strengths
- Contains 3,172 processed benchmark samples for evaluation
- Based on 351 unique egocentric videos
- Provides a clear split with 300 public development and 2,872 public test samples
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
- huggingface, author DStardust
- Collection Method
- Organized for benchmark evaluation rather than model training, as stated in the description
- Time Range
- null
- Freshness
- Last updated 2026-04-08 08:52:04
- Geography
- null