JevAny-Robotics-Train v0.4 is for non-commercial research only. It is a compilation of other datasets: the license of each record's source applies to it (see LICENSE.md), and several sources are licensed CC BY-NC-SA 4.0 or CC BY-NC 4.0. Access is granted automatically once you agree.

JevAny-Robotics-Train v0.4

The training set of SimpleJev/JevAny-Robotics-4B-LoRA v0.4: judgement questions about robot manipulation (task progress, 1–5 rollout scores, success, failure mode, plan checking, failure localisation, trajectory preference) over real and simulated robot images and videos, plus general replay questions. It is built only from the train side of public datasets and decontaminated against RoboPulse, RoboPulse++, RoboRewardBench and JevAny-Robotics-Public.

Size: 298,222 questions (train 286,795, calibration 2,279, development 9,148) and 704,700 images and video clips (50.8 GB in 21 tar shards).

Identical to the model's training set except for VLAC-Cut. VLAC-Cut's 2,650 train and val rows and the 342-row RoboPulse++ threshold set are withheld, because the VLAC-Cut repository grants no license. vlaccut_rebuild/ recreates them from your own copy of VLAC-Cut. Every other row is byte-identical to what the model was trained on, and after the rebuild every file is.

Non-commercial use only. This dataset compiles other datasets, and the license of each record's source (_meta.source, _meta.origin; table below) applies to the material derived from it. Several sources are non-commercial: airoa-moma, AgiBot World 2026 and ScienceQA are CC BY-NC-SA 4.0, and RH20T scenes 6–10 are CC BY-NC 4.0. Our own annotations (question wording, constructed options and labels) are released under CC BY-NC-SA 4.0. See LICENSE.md.

Download

Accept the terms on this page, log in (hf auth login), then:

hf download SimpleJev/JevAny-Robotics-Train-v0.4 --repo-type dataset --local-dir JevAny-Robotics-Train-v0.4
cd JevAny-Robotics-Train-v0.4 && bash unpack_media.sh   # 51 GB of media; then the shards can be deleted

Records refer to their media as media/<source>/<file>, relative to this folder. The folder is a JevAny suite: jevany.suite.load_split(folder, "train") checks each file against manifest.json.

File	Contents
`train.jsonl`	286,795 training questions
`calibration.jsonl`, `development.jsonl`	2,279 and 9,148 held-out questions, whose groups (task, scene or session) never appear in train
`val.jsonl`	calibration + development (11,427)
`media/media-000NN.tar`	21 uncompressed tar shards with 704,700 files; `media_files.tsv` maps every file to its shard, `media_shards.json` gives each shard's size and sha256
`unpack_media.sh`	unpacks the shards and checks that every referenced file exists
`manifest.json`	JevAny suite manifest: sha256 and record count of each file
`stats.json`, `samples.md`	counts by source, family and target type; one example per family and file
`build_manifest.json`, `verify.json`	how the set was assembled; the checks run on this release
`vlaccut_rebuild/`	script, keyframe list and our instruction translations for rebuilding the VLAC-Cut rows

Sources and licenses

Val = calibration + development. Upstream names are Hugging Face dataset repositories.

Source (`_meta.source`)	Train	Val	License	Upstream
Robo-Dopamine GRM data (`grm`)	80,752	1,592	Apache-2.0 ¹	`tanhuajie2001/Robo-Dopamine-GRM-Dataset`
RoboReward, train split (`roboreward`)	61,788	2,700	CC BY 4.0	`teetone/RoboReward`
DenseReward (`densereward`)	27,974	1,892	Apache-2.0	`densereward/DenseReward`
Guardian, RLBench (`guardian_rlbench`)	23,557	500	Apache-2.0	`paulpacaud/rlbenchfail_train_dataset`
Guardian, BridgeData V2 (`guardian_bdv2`)	15,961	948	Apache-2.0	`paulpacaud/bdv2fail_train_dataset`
ViFailback (`vifailback`)	9,566	817	MIT	`sii-rhos-ai/ViFailback-Dataset`
airoa-moma (`airoa`)	9,471	1,022	CC BY-NC-SA 4.0	`airoa-org/airoa-moma`
RoboVQA (`robovqa`)	7,997	500	CC BY 4.0	RoboVQA, via `Tianli/robovqa`
RoboFAC (`robofac`)	6,548	843	MIT	`MINT-SJTU/RoboFAC-dataset`
RoboCOIN (`robocoin`)	3,602	170	Apache-2.0	340 `RoboCOIN/*` repositories
REASSEMBLE, train split (`reassemble`)	1,845	60	CC BY 4.0	`robot-lev/reassemble`
AgiBot World 2026, RL rollouts (`agibot`)	1,680	195	CC BY-NC-SA 4.0	`agibot-world/AgiBotWorld2026`
RH20T, configurations 3–4 (`rh20t`), scenes 1–5	641	24	CC BY-SA 4.0	`robot-lev/rh20t_cfg3`, `robot-lev/rh20t_cfg4`
RH20T, configurations 3–4 (`rh20t`), scenes 6–10	639	13	CC BY-NC 4.0	same
RoboArena DataDump 2026-07-17 (`roboarena`)	1,136	70	MIT	`RoboArena/DataDump_07-17-2026`
RoboMIND, failure data (`robomind`)	969	81	Apache-2.0	`x-humanoid-robomind/RoboMIND`
VLAC-Cut, ARX data (`vlaccut`)	2,516 withheld	134 withheld	no license terms	`InternRobotics/VLAC-Cut-Benchmark`
General replay
HelpSteer3 (`helpsteer3`)	13,267	–	CC BY 4.0	`nvidia/HelpSteer3`
A-OKVQA (`aokvqa`)	4,990	–	see ²	`HuggingFaceM4/A-OKVQA`
Hermes function calling (`agent_tool`)	4,502	–	Apache-2.0	`NousResearch/hermes-function-calling-v1`
ScienceQA (`scienceqa`)	3,410	–	CC BY-NC-SA 4.0 ³	`derek-thomas/ScienceQA`
CommonsenseQA (`commonsense`)	2,994	–	MIT	`tau/commonsense_qa`
QASC (`qasc`)	2,662	–	CC BY 4.0	`allenai/qasc`
VideoFeedback (`video_feedback`)	500	–	Apache-2.0	`TIGER-Lab/VideoFeedback`
ARC-Challenge (`arc`)	344	–	CC BY-SA 4.0	`allenai/ai2_arc`

¹ GRM's images come from AgiBot World, AgileX, DROID, EgoDex, Galaxea, LIBERO, Pika, RoboCasa, RoboTwin and dual-arm Franka data; those datasets' own terms may also apply, and some are non-commercial. The upstream repository asks you to accept its terms. ² The A-OKVQA repository (allenai/aokvqa) is Apache-2.0; its images are COCO 2017 images, whose use follows the Flickr terms of use. ³ The original ScienceQA release is CC BY-NC-SA 4.0; the mirror we used lists CC BY-SA 4.0, and so does _meta.license in those rows. Where _meta.license (kept byte-identical to the training set) differs from this table, follow the table.

Each row's _meta.source and _meta.license make it easy to drop sources and retrain. Dropping the non-commercial sources does not make the rest safe for commercial use: GRM's images come from datasets with their own terms.

What is asked

Target type	Train	Calibration	Development	Question
`robopulse`	88,751	281	1,711	is the AFTER observation closer to the goal than the BEFORE one, given start and end references (RoboPulse style)
`rrb`	44,917	392	1,608	1–5 progress score of a rollout video (RoboRewardBench style)
`replay`	32,669	–	–	general text, image and tool-use judgement
`jj_success`	32,433	361	1,246	did the step or trajectory succeed
`jj_mode_plan`	30,933	396	1,455	failure mode; is a plan correct; what is wrong with a plan
`robopulsepp`	26,873	525	1,869	did progress rise or fall between consecutive observations, 1–3 camera views (RoboPulse++ style)
`jj_pref`	19,533	144	696	which of two trajectories is better
`jj_loc`	10,686	180	563	at which step or frame the failure happens

jj_* follow the task types of JevAny-Robotics-Public. Questions are yes/no (noul, 170,202), multiple choice (choice, 85,105) or 1–5 scores (score, 47,417). Labels come from the source's human annotation (human_expert, 62,203; human_multi, 13,767), from how the question was built (construction, 167,099: for example time order, or a plan we perturbed), from RoboReward's AI-assisted labels (ai_assisted, 55,079) or from an AI check of human failure labels (ai_verified, 74).

By family

Family	Train	Calibration	Development	Label source
`grm_rp_pair`	80,752	199	1,393	construction
`roboreward_train_progress`	44,917	392	1,608	ai_assisted, human_expert
`densereward_pp_step`	19,975	521	971	construction
`roboreward_train_traj_pref`	16,871	128	572	ai_assisted, human_expert
`replay_helpsteer3`	13,267	–	–	human_multi
`guardian_rlbench_exec_success`	10,846	100	100	construction
`densereward_rp_pair`	7,999	82	318	construction
`robovqa_train_exec_success`	7,997	99	401	human_expert
`airoa_rpp_pair`	6,778	–	892	human_expert, construction
`guardian_bdv2_exec_success`	6,276	54	313	construction
`guardian_rlbench_exec_failure_mode`	5,400	38	37	construction
`replay_aokvqa`	4,990	–	–	human_expert
`guardian_rlbench_plan_correct`	4,876	76	74	construction
`replay_agent_tool`	4,502	–	–	human_expert
`guardian_bdv2_plan_correct`	4,315	48	242	construction
`robofac_failure_step`	3,768	–	343	construction
`vifailback_failure_step`	3,460	90	110	human_expert
`vifailback_failure_frame`	3,458	90	110	human_expert
`replay_scienceqa`	3,410	–	–	human_expert
`guardian_bdv2_exec_failure_mode`	3,195	23	122	construction
`replay_commonsense`	2,994	–	–	human_expert
`replay_qasc`	2,662	–	–	human_expert
`guardian_rlbench_plan_failure_mode`	2,435	38	37	construction
`guardian_bdv2_plan_failure_mode`	2,175	24	122	construction
`robocoin_plan_correct`	1,871	6	82	human_expert, construction
`vifailback_failure_type`	1,500	137	158	human_expert
`robofac_failure_type`	1,500	–	300	construction
`agibot_exec_failure_mode`	1,471	–	156	human_expert
`robofac_traj_success`	1,280	42	158	construction
`roboarena_traj_pref`	1,136	10	60	human_expert
`vifailback_success`	1,084	47	65	human_expert
`airoa_traj_success`	1,042	–	52	human_expert
`airoa_exec_success`	1,040	–	52	human_expert
`robomind_exec_failure_mode`	969	–	81	human_expert
`robocoin_plan_failure_mode`	944	4	40	construction
`reassemble_traj_success`	738	6	18	human_expert
`reassemble_exec_success`	738	5	19	human_expert
`robocoin_traj_success`	526	–	26	human_expert, construction
`airoa_traj_pref`	521	–	26	human_expert
`replay_video_feedback`	500	–	–	human_multi
`rh20t_traj_success`	382	4	8	human_expert
`rh20t_exec_success`	382	4	8	human_expert
`reassemble_traj_pref`	369	1	11	human_expert
`replay_arc`	344	–	–	human_expert
`rh20t_traj_pref`	324	2	5	human_expert
`robocoin_traj_pref`	261	3	9	construction
`rh20t_exec_failure_mode`	192	2	4	human_expert
`agibot_traj_success`	102	–	26	human_expert, construction
`vifailback_rpp_pair`	64	4	6	ai_verified
`airoa_plan_correct`	60	–	–	human_expert, construction
`agibot_rpp_pair`	56	–	–	human_expert
`agibot_traj_pref`	51	–	13	construction
`airoa_plan_failure_mode`	30	–	–	construction

GRM by dataset family (train): AgiBot 10,941, AgileX 10,941, DROID 10,941, EgoDex 10,941, RoboTwin 10,935, Galaxea 8,000, RoboCasa 8,000, LIBERO 4,803, Pika 2,910, dual-arm Franka 2,340.

Record format

One JSON object per line, in JevAny's System One format. state is what the judge is shown, questions.decision the question and its label, media the attached images or videos, and _meta where the record comes from. A RoboRewardBench-style record, with the instructions and criteria shortened:

{"state": {"task": "Pick fruit", "video": "the rollout (first to last frame)"},
 "questions": {"decision": {"type": "score",
   "instructions": "You are watching one rollout of a robot attempting a manipulation task. ... Rate the robot's end-of-episode progress on the task on a 1-5 scale ...",
   "criteria": ["1 - No success: ...", "2 - Minimal progress: ...", "3 - Partial completion: ...", "4 - Near completion: ...", "5 - Perfect completion: ..."],
   "label": 0, "src": "roboreward_train_progress"}},
 "media": [{"type": "video", "uri": "media/roboreward/99584eef89e6a37539f3839c.mp4"}],
 "_meta": {"id": "roboreward/train/05252", "group_id": "roboreward/berkeley_mvp_converted_externally_to_rlds/pick fruit",
   "source": "roboreward", "origin": "teetone/RoboReward@469b9af76e8539e8d2ac553b081307738fd92ca5",
   "family": "roboreward_train_progress", "label_source": "ai_assisted", "license": "CC-BY-4.0", "split": "train",
   "target_type": "rrb", "extra": {"subset": "berkeley_mvp_converted_externally_to_rlds", "reward": 1, "...": "..."},
   "n_media": 1, "approx_tokens": 192, "...": "..."}}

label is true / false for noul, the key of the right option in criteria for choice, and the score minus one for score (0 = score 1).
_meta.group_id is the unit of the train / val split (task, scene or session); _meta.origin is repository@revision.
4,502 tool-use records (agent_tool) carry a second question.

How it was built

Only the train side of each source; no item of any benchmark's test or eval split.
Decontamination against RoboPulse, RoboPulse++, RoboRewardBench and JevAny-Robotics-Public: their episodes are blocked by key, and every image and video frame (2 fps) is pHash-checked against 168,861 frames of those benchmarks. A match within Hamming distance 4, confirmed by a pixel MAD below 6/255, removes the whole episode or session.
RoboReward rows follow the official split: they share no episode with the RoboRewardBench test set, but they are not frame-disjoint from it (3,771 rows have a frame that matches a test video of the same scene).
Calibration and development are held out from train by group and were never trained on.

Known label caveats. RoboReward clips cut before the end of an episode carry labels that a careful annotator reproduces only to within ±1 (20% exactly, 80% within one point). RoboReward's RoboArena-derived scores give credit for attempting. Instruction length alone predicts a RoboReward score below 5 (AUROC 0.70 on train, 0.68 on the RoboRewardBench test), because counterfactual instructions are longer.

Rebuilding the VLAC-Cut rows

InternRobotics/VLAC-Cut-Benchmark is public, but its LICENSE is a placeholder, so we redistribute neither its frames nor its annotations. Its ARX rows (2,516 train, 33 calibration, 101 development and the 342-row rpp_threshold_dev.jsonl) are RoboPulse++-style pairs of consecutive human keyframes, including 1,258 real falls in train (363 of them with several views). To rebuild them:

hf download InternRobotics/VLAC-Cut-Benchmark splits/train.json data/train_videos.tar --repo-type dataset \
    --revision fd312e11643bfd7d747a76cad57f0330c4cc0ec0 --local-dir VLAC-Cut-Benchmark   # about 90 GB
pip install av pillow numpy
python vlaccut_rebuild/rebuild_vlaccut.py --vlac VLAC-Cut-Benchmark --workers 8          # about 16 GB of RAM

The script reads the instruction, progress values and action description of each keyframe from your copy, decodes the 7,733 frames, and writes the five full JSONL files with every row at its original line, plus the original manifest.json (the released files are kept as *.without_vlaccut.*). It then checks each file against the sha256 of the training set (vlaccut_rebuild/original_files.json). From this folder it uses only the keyframe pairs and views we chose (vlaccut_index.jsonl) and our English translations of the 688 instructions (instructions_en.json). On our machine it reproduced every file and every frame byte for byte; other FFmpeg or Pillow builds may encode some frames slightly differently, which changes no row.

Verification

Run on this folder before upload (verify.json):

JevAny load_split passes for train, calibration and development; every source is declared in manifest.json.
Every line equals the training-set line, in order, with only the VLAC-Cut lines removed.
No duplicate ids; no group shared between train and val.
The 704,700 referenced media files and the files in the 21 shards are the same set, with matching sizes and shard sha256.

Acknowledgements

This dataset exists because the authors of the datasets above released their data. Please cite them when you use the corresponding parts. The model trained on it, its evaluation and its full acknowledgements are on the model card.

Downloads last month: -

Total file size:

53 GB