Video dataset discovery: human preferences, video understanding and world-model evaluation are separate tasks

This discovery index uses the archived normalized Hugging Face listing observed 2026-10-28T09:41:22.909Z. The complete bounded inventory contains 50 entries; 6 are selected below for distinct research questions. All entries remain preserved in the original inventory. This is a page-level snapshot, not a census of the Hub. Selection indicates what to inspect next, not a tested recommendation. A video modality filter returns several different research tasks. Comparing generated video preferences needs evidence about raters, criteria, prompts and clips. Video question answering needs evidence about questions, answers and scoring. Action-conditioned simulation needs evidence about actions and temporal rollouts. These cannot be treated as interchangeable evaluation sets solely because the listing labels them video. | Dataset pointer | Listed repository revision | Declared license | Gating | Listed task categories | | --- | --- | --- | --- | --- | | [Rapidata/text-2-video-human-preferences](https://huggingface.co/datasets/Rapidata/text-2-video-human-preferences) | 2ea6aeafb8fcd2fe6063075a1a3a839e86c5961a | "apache-2.0" | false | ["text-to-video", "video-classification"] | | [nvidia/MMOU](https://huggingface.co/datasets/nvidia/MMOU) | 60fddffb699443e618148f6e3ef84bb63f039cf5 | "apache-2.0" | false | ["video-text-to-text"] | | [nyu-visionx/solaris-eval-datasets](https://huggingface.co/datasets/nyu-visionx/solaris-eval-datasets) | 136ed8683de570a9e88b1f54cab1ffbf50744297 | "apache-2.0" | false | ["image-to-video"] | | [RyanWW/DynSuperCLEVR](https://huggingface.co/datasets/RyanWW/DynSuperCLEVR) | 41b4941f1ed462a7b3159e16189c9832427fc199 | "apache-2.0" | false | ["visual-question-answering", "video-text-to-text"] | | [OmniGUI/OmniGUI](https://huggingface.co/datasets/OmniGUI/OmniGUI) | 2c9031077541186f6fb1f2514cb62446c811ffd4 | "cc-by-nc-sa-4.0" | false | ["image-text-to-text", "video-text-to-text", "audio-text-to-text"] | | [RoboCOIN/RMC-AIDA-L_plate_storage](https://huggingface.co/datasets/RoboCOIN/RMC-AIDA-L_plate_storage) | 743098b52afeb13614dcc2ed92a091af1153db99 | "apache-2.0" | "auto" | ["robotics"] | Rapidata/text-2-video-human-preferences: Listed text-to-video and video-classification tasks, with human/preferences/coherence/style/alignment tags, make this a concrete candidate for inspecting the human-rating protocol. The tags do not establish how raters worked or how any named model scored. Listed modalities: tabular, text, video. Pinned card pointer, not inspected: https://huggingface.co/datasets/Rapidata/text-2-video-human-preferences/raw/2ea6aeafb8fcd2fe6063075a1a3a839e86c5961a/README.md nvidia/MMOU: Listed video-text-to-text and long-video/reasoning/benchmark tags point toward understanding evaluation. Questions, answers, difficulty, split integrity and evaluation code still need inspection. Listed modalities: text, video, audio. Pinned card pointer, not inspected: https://huggingface.co/datasets/nvidia/MMOU/raw/60fddffb699443e618148f6e3ef84bb63f039cf5/README.md nyu-visionx/solaris-eval-datasets: Listed image-to-video task and minecraft/world-model/multi-agent/action-conditioned tags identify a distinct simulation research route. They do not establish controls, rollout lengths or transfer to ordinary generated-video evaluation. Listed modalities: video. Pinned card pointer, not inspected: https://huggingface.co/datasets/nyu-visionx/solaris-eval-datasets/raw/136ed8683de570a9e88b1f54cab1ffbf50744297/README.md RyanWW/DynSuperCLEVR: Listed visual-question-answering/video-text-to-text and physics/spatial reasoning tags identify a possible synthetic reasoning source. No physics or temporal-consistency metric was inspected. Listed modalities: image, tabular, text, video. Pinned card pointer, not inspected: https://huggingface.co/datasets/RyanWW/DynSuperCLEVR/raw/41b4941f1ed462a7b3159e16189c9832427fc199/README.md OmniGUI/OmniGUI: Listed multimodal text-generation tasks and GUI/agent tags illustrate that video modality does not imply video-generation evaluation. The noncommercial/share-alike license declaration requires full applicable terms review. Listed modalities: audio, image, video. Pinned card pointer, not inspected: https://huggingface.co/datasets/OmniGUI/OmniGUI/raw/2c9031077541186f6fb1f2514cb62446c811ffd4/README.md RoboCOIN/RMC-AIDA-L_plate_storage: Listed robotics task and auto gate illustrate a video-containing robotics pointer with a separate access step. The Apache declaration does not remove the gate or establish rights to particular recordings. Listed modalities: tabular, timeseries, video. Pinned card pointer, not inspected: https://huggingface.co/datasets/RoboCOIN/RMC-AIDA-L_plate_storage/raw/743098b52afeb13614dcc2ed92a091af1153db99/README.md For generated-video preference research, inspect Rapidata first for actual rating dimensions, annotator protocol, model versions, prompts and coverage; the listed model-name tags are not measured results. For understanding, MMOU and DynSuperCLEVR point to question/answer or reasoning tasks, not demonstrated generator-quality tests. Solaris is a separate action-conditioned/world-model discovery route. A robotics or GUI dataset can involve video without measuring text-to-video realism, subject consistency or instruction following. These are proposed review priorities; no candidate is endorsed as a valid benchmark. The license and gate columns are API metadata declarations only. false gating does not establish permission; auto/manual gating does not establish actual access. null is unknown, and other requires the actual terms. A repository-level declaration does not settle source-media licenses, derivative copies, annotations, identifiable-person permissions or commercial use. Cards, license text, files and rows were not fetched or inspected. Before benchmark or training use, inspect the pinned card and applicable full terms, identify the actual files/splits and their hashes, check provenance, leakage and overlap with training data, define the task and scoring protocol, and establish whether labels, prompts, reference media and evaluation code support that task. For video, check clip duration, sampling/FPS, frame geometry, temporal alignment and audio requirements from actual data. Record exact revisions, split/sample IDs, preprocessing and metric/code versions; these are proposed evidence checks, not a reproduced run. The listed repository revision is a discovery anchor, not a checksum of inspected dataset files or an executed loader. Modality and task tags do not establish that rows contain decodable video/images, paired edit instructions, identity references, human ratings or a valid held-out evaluation. Downloads and likes are snapshot counters, not proof of quality, independence or external adoption. Evidence: metadata_checked / not_run / unreviewed. The retained object is a deterministic metadata projection, not the original raw API response or linked card text. No media, dataset rows, model, loader or benchmark was executed. Reviewers share the founder operator; no independent outside attestation is claimed.

{
  "citations": [
    {
      "url": "https://huggingface.co/api/datasets?cursor=eyIkb3IiOlt7ImRvd25sb2FkcyI6MjA0MywiX2lkIjp7IiRndCI6IjZhNjIyNjg0YzNlMDUwNTIxMjI2MjhjMSJ9fSx7ImRvd25sb2FkcyI6eyIkbHQiOjIwNDN9fSx7ImRvd25sb2FkcyI6bnVsbH1dfQ%3D%3D&direction=-1&expand=sha&expand=gated&expand=lastModified&expand=createdAt&expand=downloads&expand=likes&expand=tags&filter=modality%3Avideo&limit=50&sort=downloads",
      "note": "Archived normalized listing metadata; exact inventory locators/hashes in source pins. Linked cards and rows were not inspected."
    }
  ]
}