EgoCoT-Bench

Project Page | Paper

EgoCoT-Bench is a benchmark for evaluating grounded and verifiable reasoning in egocentric videos.

The benchmark focuses on operation-centric first-person video understanding, including hand-object interaction, object state changes, temporal retrospection, short-horizon anticipation, local causal reasoning, and grounded rationale evaluation.

Note. This is the official clean public release of EgoCoT-Bench. The benchmark contents and public splits remain unchanged, with 300 public development samples and 2,872 public test samples. The media/ directory hosts only 63 self-recorded videos for which the authors have release rights. For third-party samples, this release provides annotations, source/media references, preprocessing metadata, a reconstruction manifest, and reconstruction scripts; users must obtain the original media from the corresponding dataset providers.

Overview

This repository contains the public release of EgoCoT-Bench. The release is organized for benchmark evaluation and qualitative analysis rather than model training.

• Total samples in the processed benchmark: 3,172
• Total unique videos in the processed benchmark: 351
• Public development samples: 300
• Public development unique videos: 193
• Public test samples: 2,872
• Public test unique videos: 351

The official hidden test labels are not included in this public repository.

Release Scope

This repository contains the public benchmark splits and annotations for EgoCoT-Bench. The public splits remain:

• 300 public development samples
• 2,872 public test samples

The public development split may include answers and public rationale annotations for inspection, method development, and reproducibility. The public test split omits official answers and private rationale annotations to avoid label leakage.

For media release, this repository follows a clean-release policy:

• For self-recorded samples, this repository provides 63 released self-recorded videos under the EgoCoT-Bench license.
• For third-party samples derived from Ego4D, EPIC-KITCHENS, Charades-Ego, MECCANO, HD-EPIC, or other external datasets, this repository provides benchmark annotations, source identifiers, media references, preprocessing metadata, reconstruction metadata, and evidence metadata where released.
• This repository does not redistribute third-party raw videos or derived/scaled clips.
• Users must obtain third-party media from the original dataset providers and comply with the corresponding source dataset licenses, access agreements, and redistribution policies.
• Licensed users may reconstruct the processed/scaled benchmark clips locally using the released reconstruction manifest and scripts.

In short, the public benchmark contains all 3,172 QA samples, but the media/ directory only hosts the 63 self-recorded videos for which the EgoCoT-Bench authors have release rights.

Files

• data/: public split files and related benchmark annotations. This includes the 300-sample public development split and the 2,872-sample public test split.
• media/: released self-recorded media only. Third-party raw videos or derived/scaled clips are not redistributed in this directory.
• manifests/media_reconstruction.csv: reconstruction metadata for mapping third-party benchmark samples to locally reconstructed processed/scaled clips.
• scripts/reconstruct_media.py: local reconstruction utility for licensed users who have obtained the original third-party videos from the corresponding dataset providers. This script does not download, upload, or redistribute third-party media.
• LICENSE: primary license notice for EgoCoT-Bench annotations, benchmark metadata, and self-recorded media where provided.
• LICENSES/Apache-2.0.txt: license notice for evaluation scripts, data-loading utilities, and leaderboard-related code, where included.
• LICENSES/CC-BY-NC-4.0.txt: license notice for EgoCoT-Bench annotations, benchmark metadata, and self-recorded media where provided.
• THIRD_PARTY_NOTICES.md: source-wise notices for third-party datasets and media access requirements, where provided.

For third-party samples, released media paths or media references should be interpreted as benchmark references or expected local reconstruction paths. They do not indicate that the corresponding third-party video file is hosted in this repository.

Data Fields

Each sample may include the following fields, depending on the split and release setting:

• qid: unique sample identifier
• question: question text
• choices: multiple-choice options
• answer: answer label, released for public development examples only
• cot: public rationale annotation, released for public development examples only where available
• media_id: video or media identifier
• source: for self-recorded samples, a relative path to released media in this repository; for third-party samples, a benchmark media reference or expected local reconstruction path rather than a hosted media file
• big_category_en: high-level task group
• sub_category_en: fine-grained subtask
• type: original task type label
• evidence: released evidence-related metadata, where available

For third-party samples, fields such as source and media_id may preserve the processed benchmark media identifiers used by the original evaluation pipeline. These fields are provided for reproducibility and local reconstruction only; the corresponding third-party raw videos or derived/scaled clips are not hosted in this repository. Public test examples do not include official answers or private rationale annotations.

License

EgoCoT-Bench uses a layered licensing model.

The Hugging Face dataset license metadata, cc-by-nc-4.0, applies to EgoCoT-Bench annotations, benchmark metadata, public rationale annotations where released, task labels, released evidence metadata where available, and self-recorded media where provided.

Evaluation scripts, data-loading utilities, and leaderboard-related code are released under the Apache License 2.0, unless otherwise specified.

Third-party videos or clips derived from third-party datasets are not re-licensed by EgoCoT-Bench and are not redistributed in this clean public release. Media derived from sources such as Ego4D, EPIC-KITCHENS, Charades-Ego, MECCANO, and HD-EPIC remains governed by the original source dataset licenses, access agreements, and redistribution policies.

For third-party samples, EgoCoT-Bench releases benchmark annotations, source identifiers, temporal spans, media references, and evidence metadata where available. Users are responsible for obtaining the corresponding third-party media from the original dataset providers and ensuring that their use complies with the applicable source dataset terms.

Users should not redistribute third-party raw videos or derived clips unless such redistribution is explicitly permitted by the original source license or separately authorized by the rights holder.

By using EgoCoT-Bench, users agree to use the benchmark for research and non-commercial purposes, comply with applicable third-party media terms, and cite EgoCoT-Bench together with relevant source datasets when appropriate.

Maintenance and Versioning

This release is the official clean public release of EgoCoT-Bench. The benchmark contents, sample counts, public splits, and evaluation protocol remain unchanged. The release update clarifies the media redistribution boundary: the repository hosts only self-recorded media, while third-party samples are provided as annotations, references, preprocessing metadata, and local reconstruction metadata.

We will maintain the dataset repository and project page with versioned updates. Future corrections, metadata updates, or evaluation-protocol changes will be documented in the dataset card and repository history.

Loading Example

from datasets import load_dataset

ds = load_dataset("DStardust/EgoCoT-Bench")

dev = ds["dev"]
test = ds["test"]

print(len(dev), len(test))
print(dev[0].keys())

Project Page

Homepage: https://dstardust.github.io/EgoCoT/

Paper: EgoCoT-Bench: Benchmarking Grounded and Verifiable Operation-Centric Chain-of-Thought Reasoning for MLLMs