EgoSAT

Dataset Description

EgoSAT is an ECCV 2026 benchmark for egocentric streaming interaction understanding:

EgoSAT: A Comprehensive Benchmark of Egocentric Streaming Interaction Understanding

The benchmark evaluates vision-language models under streaming constraints, including retrospective reasoning, present-time interaction understanding, and prospective reasoning over future actions or state transitions.

This HuggingFace dataset contains the main EgoSAT release candidate: ground-truth annotations, official shuffled multiple-choice files, official effective querysets, release metadata, checksums, and sanitized SFT manifests. It does not contain raw Ego4D videos. Users must obtain Ego4D access and license approval separately before running video inference, evaluation, or training.

Related resources:

• Paper: https://arxiv.org/abs/2606.24422
• Code: https://github.com/leiyj23/EgoSAT
• Project page: https://leiyj23.github.io/EgoSAT/
• Main dataset repo: https://huggingface.co/datasets/YijiaLeithu/EgoSAT
• ROI cache dataset repo: https://huggingface.co/datasets/YijiaLeithu/EgoSAT-ROI-Cache

DOI, proceedings volume, and page numbers will be added after proceedings publication.

Dataset Structure

.
  README.md
  egosat/
    gt/
    mcq_shuffled/
    effective_querysets/
    metadata/
  sft/

• egosat/gt/: released ground-truth annotations for the public task names.
• egosat/mcq_shuffled/: official shuffled MCQ candidate files for the paper main-table seed.
• egosat/effective_querysets/: official evaluation subsets used by the runner and scorer.
• egosat/metadata/: release manifests, checksums, task statistics, video UID list, and GT source mapping.
• sft/: sanitized SFT manifests and statistics.

Some files use the .jsonl extension for historical compatibility, but many per-video or per-clip files store one full JSON object rather than one JSON object per line. For those files, use json.load() or json.loads(full_text) instead of line-by-line JSONL parsing unless a subdirectory README documents otherwise.

Tasks

Ground-truth task directories:

now_narration
now_state_switch
sh_pred
sh_rtrv
ms_pred
ms_rtrv

Multiple-choice task directories:

now_narration_action
sh_pred
sh_rtrv
ms_pred
ms_rtrv

now_narration and now_state_switch public GT are derived from interaction GT sources. The mapping is recorded in egosat/metadata/gt_source_mapping.json:

now_narration    -> gt/now_narration_interaction
now_state_switch -> gt/now_state_switch_interaction

The optional alternate MCQ seed mcq_shuffled_second_seed20260506 is not included in this main release candidate.

Field-Level Schema

GT Files

GT files are per-video or per-clip JSON objects. Sampled files contain top-level fields such as:

• dataset: dataset identifier.
• video_uid: Ego4D video UID.
• video_metadata: clip/video metadata used for alignment.
• task: EgoSAT task name.
• params: task generation parameters such as lookback or horizon settings.
• num_samples: number of sample records in the file.
• samples: task samples for this video or clip.
• stats: per-file statistics.
• source_task: present for derived public tasks such as now_narration and now_state_switch.

Common samples[*] fields include:

• idx: sample index within the file.
• t_eval and t_eval_rel: evaluation timestamp, absolute and relative when available.
• lookback_sec, horizon_sec, window_start_sec, window_end_sec: temporal window metadata.
• gt_has_action, gt_action_uid, gt_segment_start, gt_segment_end: linked GT action metadata.
• gt_verb, gt_noun, noun_source, gt_text_narr, gt_text_canonical: action and narration labels.
• prompt: natural-language prompt for tasks that store prompts in the payload.
• region, visible_interaction, is_rejected, reject_reason: present-time interaction filtering fields.
• state_order, surprise_label, branchiness_label, state: state and transition labels for prediction/state-heavy tasks.
• anchor_group_id, anchor_idx, group_size, step_idx, lead_sec or lag_sec: multi-step prediction/retrieval metadata.

Task-specific files may add fields such as gap_source, gap_block_start, gap_block_end, sur_ctx_actions, or other derivation traces.

MCQ Files

MCQ files are single JSON objects despite the .jsonl suffix. Sampled files contain top-level fields such as:

• task: MCQ task name.
• split: split name.
• video_uid: Ego4D video UID.
• clip_id: EgoSAT clip or interval id.
• generator: MCQ generation metadata.
• num_samples: number of MCQ samples.
• samples: shuffled MCQ records.
• shuffled_meta: file-level shuffle metadata.

Common samples[*] fields include:

• idx: sample index.
• state: optional state label used by state/action variants.
• t_eval: evaluation timestamp.
• gt: ground-truth answer source.
• prev, future, absurd: distractor source categories.
• options: shuffled answer options, usually four choices.
• option_sources: source label for each option.
• answer: answer letter or text label.
• answer_idx: zero-based index into options.
• shuffle_meta: sample-level shuffle metadata.

The sampled official MCQ payloads do not consistently store a prompt field. Evaluation code or loaders should construct the user-facing prompt from task context, timestamps, and options when needed.

Effective Querysets

Effective querysets are the official input subset for the EgoSAT runner and scorer. They are organized by task, flavor, split, and state-switch position setting where applicable.

Top-level fields include:

• dataset, video_uid, video_metadata, task, params, num_samples, samples, and stats.
• task_name, helper, source_task, release_task, release_flavor, and release_split.
• release_sspos and release_sspos_raw for state-switch querysets.

Common samples[*] fields include:

• idx, t_eval, t_eval_rel, and optional t_target_sec.
• lookback_sec, horizon_sec, window_start_sec, and window_end_sec.
• prompt: evaluation prompt when present in the queryset payload.
• GT fields such as gt_verb, gt_noun, gt_text_narr, gt_text_canonical, and GT segment timestamps.
• mcq: nested MCQ payload with options, option_sources, answer, answer_idx, and shuffle_meta for candidate-based querysets.
• State-switch fields such as state_switch_pair_type, state_switch_role, state_switch_param, state_switch_neighbor_dt, state_switch_pair_id, and state_switch_trace.
• Multi-step fields such as anchor_group_id, anchor_idx, group_size, step_idx, lead_sec, lag_sec, and query_ref_sec.

Some downstream experiments may attach memory or context fields outside this release candidate. Loaders should treat task-specific fields as optional unless required by the selected evaluator.

SFT Manifests

The included SFT manifests are JSONL files:

sft/train_manifest_mixed5_cand.sanitized.jsonl
sft/train_manifest_mixed7_stateheavy.sanitized.jsonl

Each line is one training row. Sampled rows contain:

• dataset: source dataset identifier.
• video_uid: Ego4D video UID.
• task: EgoSAT task name.
• sft_task: SFT mixture name.
• params: task parameters.
• video_metadata: metadata needed to locate the aligned video interval.
• sample: source EgoSAT sample, including sample.prompt and task labels.
• target_text: supervised target text.
• assets: external asset requirements.

assets includes video_source, requires_ego4d_video, requires_roi_cache, and roi_cache_key. mixed5_cand and mixed7_stateheavy are two independent SFT manifests, not stage 1 and stage 2. Actual training still requires user-provided Ego4D videos. ROI-aware training also requires the separate ROI cache dataset.

Empty Payload Files

Some per-video/per-task files intentionally contain an empty samples list. In this release candidate, a text scan found 34 MCQ files and 204 effective-queryset files with samples: [].

These files are retained to preserve per-video file alignment, task layout, and release structure after filtering, alignment, or effective-queryset construction. Loaders and evaluators should skip files whose samples list is empty. Empty payload files do not indicate missing raw videos or corrupted JSON.

Ego4D Videos

Raw Ego4D RGB videos are not redistributed in this dataset. Users must separately obtain Ego4D access and comply with the Ego4D license and terms of use. This dataset does not grant any rights to the original Ego4D videos.

Configure a local Ego4D video root before running workflows that read video frames, for example:

export EGO4D_VIDEO_ROOT=/path/to/Ego4D/videos

License and Usage Terms

Dataset license: other.

EgoSAT annotations, querysets, SFT manifests, and ROI-cache references are released for research use. Raw Ego4D videos are not redistributed in this dataset. Users must separately obtain Ego4D access and comply with the Ego4D license and terms of use. This dataset does not grant any rights to the original Ego4D videos.

The code repository is released separately under the MIT License. Users are responsible for ensuring that their use complies with the licenses and terms of Ego4D and EgoSAT.

Citation

If you use EgoSAT, cite the benchmark paper:

@inproceedings{lei2026egosat,
  title={EgoSAT: A Comprehensive Benchmark of Egocentric Streaming Interaction Understanding},
  author={Lei, Yijia and Li, Jinzhao and Zhang, Yichi and Hua, Jiacheng and Li, Yin and Liu, Miao},
  booktitle={European Conference on Computer Vision},
  year={2026}
}

The arXiv preprint can be cited as:

@article{lei2026egosat_arxiv,
  title={EgoSAT: A Comprehensive Benchmark of Egocentric Streaming Interaction Understanding},
  author={Lei, Yijia and Li, Jinzhao and Zhang, Yichi and Hua, Jiacheng and Li, Yin and Liu, Miao},
  journal={arXiv preprint arXiv:2606.24422},
  year={2026}
}

Responsible Use

Use this dataset only for research and evaluation with properly licensed Ego4D video access. Do not redistribute raw Ego4D videos. Do not attempt to reconstruct identities or infer private attributes from source videos. Do not treat benchmark outputs as safety-critical decisions. Public reports should identify the exact EgoSAT release candidate, model, prediction flavor, and effective queryset subset used for evaluation.