datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
HumanGenTarsier2-Recap-585K
Dataset Card for Tarsier2-Recap-585K
Introduction
✨Tarsier2-Recap-585K✨ consists of 585K distinct video clips, lasting for 1972 hours in total, from open-source datasets (e.g. VATEX, TGIF, LSMDC, etc.) and each one with a detailed video description annotated by Tarsier2-7B, which beats GPT-4o in generating detailed and accurate video descriptions for video clips of 5~20 seconds (See the DREAM-1K Leaderboard). Experiments demonstrate its effectiveness in enhancing the… See the full description on the dataset page: https://huggingface.co/datasets/omni-research/Tarsier2-Recap-585K.epstractor-raw
Epstractor: Epstein Archives Dataset
A comprehensive archive of documents, images, audio, and video files from multiple Epstein-related releases, including estate records and Department of Justice materials obtained through FOIA requests.
Dataset Description
This dataset contains 59,420 files totaling 115.23 GB from three major document releases, plus 2 large videos (40GB) available via a separate config:
Epstein Estate 2025-09: 5 files, 0.09 GB
Epstein Estate 2025-11:… See the full description on the dataset page: https://huggingface.co/datasets/public-records-research/epstractor-raw.Physics-IQ-Verified
Physics-IQ Verified Dataset
This repository hosts the Physics-IQ Verified benchmark data for evaluating physical understanding in generative video models.
Physics-IQ Verified is derived from the original Physics-IQ benchmark dataset.
Original Physics-IQ
Paper: Do generative video models understand physical principles?
Repository: Code | Dataset in Google Cloud
Physics-IQ Verified (Recommended)
Paper: Physics-IQ Verified
Repository: Code | Dataset: Here in this repo :)
We… See the full description on the dataset page: https://huggingface.co/datasets/Anates-Labs-Research/Physics-IQ-Verified.HDTF
HDTF Dataset for Talking Face Generation
This dataset is a curated and preprocessed version of HDTF (High-Definition Talking Face), prepared for tasks such as talking-head generation, video captioning, and multimodal avatar synthesis.
📁 File Overview
File Name
Description
videos.zip
Original 400 full-length HDTF videos
clips.zip
Videos split into 81-frame clips, each representing a short temporal unit
audios.zip
Audio embeddings extracted from original… See the full description on the dataset page: https://huggingface.co/datasets/global-optima-research/HDTF.QIVD
QIVD: Qualcomm Interactive Video Dataset
A collection of 2,900 video clips paired with visual question-answer annotations.
Each clip is associated with exactly one question drawn from one of 13 fine-grained QA categories,
a full-sentence answer, a concise short answer, and a timestamp pinpointing the relevant moment in the video.
Overview
QIVD is a dataset and benchmark for online, situated audio-visual question answering. Unlike existing video QA benchmarks… See the full description on the dataset page: https://huggingface.co/datasets/Qualcomm-AI-Research/QIVD.psi0-g1-sneaker-205ep-v2-source
Psi0 G1 Sneaker-in-Box — 205 episodes (v2 canonical source)
⚠️ Do not use this dataset directly for training. This is the canonical immutable union of the v1 and v2 collections, kept as a source of truth for reproducibility. For v2 fine-tuning use psi0-g1-sneaker-199ep-v2; for held-out evaluation use psi0-g1-sneaker-6ep-v2-eval. Together these two derivatives reconstruct this canonical dataset exactly: 199 + 6 = 205.
205 teleoperated episodes of a Unitree G1 humanoid (with Inspire… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/psi0-g1-sneaker-205ep-v2-source.gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-5
Grab-Bottle (right hand) - zero-wandering, smoothing half-width 5
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task, curated to
remove non-purposeful arm motion ("wandering") so a trained policy does not learn
unstable behavior. Derived from the
105-episode source teleoperation set
by running the kinematic zero-wandering pipeline at smoothing half-width 5;
each surviving clean sub-segment is emitted as its own episode (no action chunk ever
spans a cut).
This is… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-5.gr00t-g1-grab-bottle-right-hand-105ep-v1
GR00T G1 Grab-Bottle — right hand, 105 episodes (source teleoperation set)
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. This is the
base source teleoperation set of the grab-bottle lineage — 105 episodes of
real-robot demonstration, manually cleaned at the episode level (bad runs discarded),
with no automated frame-level curation applied.
It is the dataset the curated variants are derived from: the
zero-wandering-smooth-radius-{5,10,15,20} sets and the… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-105ep-v1.gr00t-g1-grab-bottle-right-hand-speedup-3mm-v1
Grab-Bottle (right hand) - DP-resampled, 3 mm/frame wrist target
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. Merges two
teleoperation datasets and applies Dynamic Programming (DP) resampling on the
right-wrist XYZ trajectory to normalise the action density across all episodes. It is
the training data for the
v7 GR00T N1.7 fine-tune
a speedup-resampling reference baseline (trained 2026-06-25, checkpoint-20000
only).
The key idea: instead of keeping… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-speedup-3mm-v1.DREAM-1K
DREAM-1K
DREAM-1K (Description
with Rich Events, Actions, and Motions) is a challenging video description benchmark. It contains a collection of 1,000 short (around 10 seconds) video clips with diverse complexities from five different origins: live-action movies, animated movies, stock videos, long YouTube videos, and TikTok-style short videos. We provide a fine-grained manual annotation for each video.
Bellow is the dataset statistics:
gr00t-g1-grab-bottle-right-hand-radius-20-merged
Grab-Bottle (right hand) - radius-20 merged
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. This is a
merge of two source teleoperation datasets, curated with the zero-wandering
pipeline at smoothing half-width 20 and safe_frames 50, then renumbered into
one contiguous set of 502 episodes. It is the largest and best-curated set of the
grab-bottle lineage and the training data for the
v6 GR00T N1.7 fine-tune.
Each surviving clean sub-segment is emitted as its… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-radius-20-merged.gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-15
Grab-Bottle (right hand) - zero-wandering, smoothing half-width 15
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task, curated to
remove non-purposeful arm motion ("wandering") so a trained policy does not learn
unstable behavior. Derived from the
105-episode source teleoperation set
by running the kinematic zero-wandering pipeline at smoothing half-width 15;
each surviving clean sub-segment is emitted as its own episode (no action chunk ever
spans a cut).
This… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-15.gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-10
Grab-Bottle (right hand) - zero-wandering, smoothing half-width 10
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task, curated to
remove non-purposeful arm motion ("wandering") so a trained policy does not learn
unstable behavior. Derived from the
105-episode source teleoperation set
by running the kinematic zero-wandering pipeline at smoothing half-width 10;
each surviving clean sub-segment is emitted as its own episode (no action chunk ever
spans a cut).
This… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-10.gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1
Grab-Bottle (right hand) - DP-resampled 3 mm/frame + cycle removal
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. Extends the
DP-resampled variant
by automatically detecting and removing return cycles in wrist-space - segments where
the arm travels away from a point and then returns to (approximately) the same position.
Such cycles teach the policy to backtrack rather than commit to a goal-directed motion.
No fine-tune has been produced from this variant… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-speedup-3mm-cycle-removed-v1.gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-20
Grab-Bottle (right hand) - zero-wandering, smoothing half-width 20
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task, curated to
remove non-purposeful arm motion ("wandering") so a trained policy does not learn
unstable behavior. Derived from the
105-episode source teleoperation set
by running the kinematic zero-wandering pipeline at smoothing half-width 20;
each surviving clean sub-segment is emitted as its own episode (no action chunk ever
spans a cut).
This… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-zero-wandering-smooth-radius-20.gr00t-g1-grab-bottle-right-hand-v2
Grab-Bottle (right hand) - v2 (distance-to-goal wandering removal)
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. Curated from
the 105-episode source set
by removing non-purposeful back-and-forth arm motion ("wandering") via a
distance-to-goal method, then splitting each episode into clean sub-episodes so
GR00T's 40-step action window never spans a cut. It is the training data for the
v2 GR00T N1.7 fine-tune
the current validated production champion (11/12… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v2.pick-and-place-banana_v2This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "bi_piper_follower",
"total_episodes": 996,
"total_frames": 403881,
"total_tasks": 2,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 500,
"fps": 30,
"splits": {
"train": "0:996"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/NONHUMAN-RESEARCH/pick-and-place-banana_v2.libero_rollouts_tests_100epgr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3
GR00T G1 Grab-Bottle — right hand, worst-positions, 115 episodes
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. This is a
hard-pose teleoperation set — 115 episodes collected deliberately at the bottle
positions where prior fine-tuned policies performed worst (the model would approach
the bottle but stop ~10 cm short of contact and fail to grasp). It was recorded to fill
the policy's reach-distance distribution gap. Manually cleaned at the episode level; no… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-worst-positions-empty-115ep-v3.manus-egocentric-sample
manus-egocentric-sample
Egocentric video dataset with Manus glove hand tracking data, converted to LeRobot v3.0 format.
Dataset Description
This dataset contains egocentric (first-person view) recordings of human hands performing various manipulation tasks, captured with:
Manus Metagloves: High-precision finger tracking (~70Hz)
OAK-D Camera: RGB video (1920x1080, 30fps) + Depth (640x400, 30fps)
IMU: Accelerometer and gyroscope data
Dataset Statistics… See the full description on the dataset page: https://huggingface.co/datasets/OpenGraphLabs-Research/manus-egocentric-sample.hanoi-teleoptblock4This dataset was created using LeRobot.
Dataset Description
Bimanual HandUMI demonstration data recorded in LeRobot v3 format. Task: put the t block in the puzzle
Homepage: https://github.com/murobotics-ai/handumi-sw
Paper: [More Information Needed]
License: other
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "handumi_raw",
"total_episodes": 122,
"total_frames": 32817,
"total_tasks": 1,
"chunks_size":… See the full description on the dataset page: https://huggingface.co/datasets/NONHUMAN-RESEARCH/tblock4.gr00t-g1-grab-bottle-right-hand-v11
Grab-Bottle (right hand) - v11 (Break Down + Speed Up: v1 wandering detection then DP resampling)
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. Combines the
v1 wandering-detection method ("Break Down") with DP frame resampling ("Speed
Up") on the merge of the two source teleoperation sets. Break Down runs first to
remove wandering blocks and split into sub-segments; DP resampling is then applied
independently to each sub-segment using a dynamic per-frame… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v11.psi0-g1-sneaker-94ep-v1
Psi0 G1 Sneaker-in-Box — 94 episodes (v1)
94 teleoperated episodes of a Unitree G1 humanoid (with Inspire RH56DFTP hands) picking up a sneaker and placing it inside a box. Right-hand only, single-handed task. Used to fine-tune the v1 Psi0 checkpoint at the CloudWalk Robotics Lab (CW-RL), 2026-04. To our knowledge, the v1 fine-tune was the first publicly reported external reproduction of the Psi0 paper's full pipeline (collect → finetune → real-robot deploy).
Quick facts… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/psi0-g1-sneaker-94ep-v1.psi0-g1-he-pickplace-9x40ep
psi0-g1-he-pickplace-9x40ep
9 curated pick-and-place tasks from the Humanoid-Everyday (HE) dataset, re-targeted from the Unitree G1 + Dex3-1 3-finger hands onto our G1 + Inspire RH56DFTP 6-DoF lockstep schema — so it can be mixed directly into LeRobot v2.1 fine-tunes of Psi0 (or any policy trained against the Inspire convention) without further preprocessing.
v1.3 update (2026-05-15) — fixes a RH-only polarity inversion that v1.0/v1.1/v1.2 all carried silently. The Dex3-1… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/psi0-g1-he-pickplace-9x40ep.gr00t-g1-grab-bottle-right-hand-v10
Grab-Bottle (right hand) - v10 (Break Down: v1 wandering detection on the merged set)
LeRobot v2.1 dataset for the Unitree G1 right-hand bottle-grab task. Applies the
v1 wandering-detection method ("Break Down") to the merge of the two source
teleoperation sets: rising segments of the wrist's distance-to-goal curve are detected
and the corresponding frames are removed, and each kept sub-segment becomes its own
episode so no GR00T action chunk spans a cut.
This is the family's… See the full description on the dataset page: https://huggingface.co/datasets/cloudwalk-research/gr00t-g1-grab-bottle-right-hand-v10.tblock-all-piper-cleanThis dataset was created using LeRobot.
Dataset Description
Curated HandUMI dataset derived locally from local/tblock-all. Tasks: put the t block in the puzzle. Removed source episodes: [50, 52, 53, 55, 56, 58, 64].
Homepage: https://github.com/murobotics-ai/handumi-sw
Paper: [More Information Needed]
License: other
What was removed, and why
This dataset is tblock-all with 7 episodes removed. Every removal has the same
cause: after retargeting to the Piper… See the full description on the dataset page: https://huggingface.co/datasets/NONHUMAN-RESEARCH/tblock-all-piper-clean.libero10-rollouts-recap-single_stagepick
pick
This dataset was generated using phosphobot.
This dataset contains a series of episodes recorded with a robot and multiple cameras. It can be directly used to train a policy using imitation learning. It's compatible with LeRobot.
To get started in robotics, get your own phospho starter pack..