Fill in the form and accept the license to request access. QURE reviews each request by hand. Please tell us who you are and how you plan to use the dataset.

This dataset is shared for evaluation under the Dataset Evaluation License (LICENSE.md). You may not redistribute it or try to identify anyone who appears in it. It must not be used for clinical decisions.

Suturing Skills Action Clips

First-person video of suturing practice on a silicone pad, cut into one-action clips with structured JSON labels. Use it to fine-tune a vision-language model to describe an action or grade one against written criteria.

This is a sample. The full dataset has 3 camera angles of every action: headset, tool close-up and room view. See Full datasets.

At a glance

	Describe (`default`)	Grade (`execution_grading`)
Who is in the video	Trainer	Trainee
Model input	Clip	Clip + trainer's criteria as text
Model output	JSON description of the action	JSON grade and reason
Original clips	105 (19.2 min)	14 (8.0 min)
Source	5 protocols, 18 step videos	1 full run of protocol 1849 (15.2 min)
Train / validation	468 / 27	66 / 3

Video: 720×540, 10 fps, H.264 MP4, no audio. Language: English. Contact: QURE Industries, Inc., shar@qurexr.com

Labels

Describe: Drafted by NVIDIA Cosmos 3 Nano, then every label corrected against the video in human review.
Grade: Evaluations generated by NVIDIA Cosmos 3 Nano, then validated against the video frame by frame, with incorrect results corrected under human review.

Quick start

Request access above, log in, then:

from datasets import load_dataset

describe = load_dataset("qure-ai/suturing-skills-action-clips")
grade = load_dataset("qure-ai/suturing-skills-action-clips", "execution_grading")

sample = grade["train"][0]
sample["video"]                     # decoding needs torchcodec
sample["prompt"], sample["answer"]

Trainer	How to use it
NVIDIA Cosmos 3 Reasoner SFT	`cosmos/<split>/meta.json` or `execution_grading/cosmos/<split>/meta.json`; repo folder as data root, `video_0` as media field
TRL, LLaMA-Factory, custom loop	Map `file_name`, `prompt` and `answer` to your chat format

Training has not been run on this release yet. The layouts follow each trainer's documented format.

The two tasks

Describe an action

Target for 1849_2_a03:

{"title": "Take the tissue forceps out of the case",
 "step_context": "The user pulls the tissue forceps out of its strap in the open instrument case and lays it in front of the case.",
 "detailed_steps": ["Pinch the tissue forceps and slide it out of its strap",
                    "Lay the tissue forceps in front of the case"],
 "objects_expected": ["Tissue forceps", "Instrument case"],
 "hand_pose_expectation": {"left": "open", "right": "pinching"},
 "success_criteria": ["The tissue forceps lies in front of the case."],
 "failure_modes": ["The tissue forceps stays in its strap or is dropped."]}

Grade an action

The prompt holds the trainer's success criteria and failure modes. The trainer's clip is linked in reference_video but is not part of the prompt.

Target for 1849_5_a01_trainee, graded against "Make a second stitch across the incision":

{"observed": true,
 "outcome_score": 0,
 "recovered": false,
 "verdict": "The thread is wrapped around the forceps and the short tail is pulled through with the needle holder, tying a knot on the first stitch. The needle stays off the pad, so no second pass crosses the cut.",
 "deviations": [{"why": "A knot was tied on the first stitch in place of a second stitch.", "evidence": "The thread enters the pad only at the one spot of the first stitch, with the free length running off the pad to the forceps."}]}

The trainee run

Result
Actions graded	14: 9 done, 2 recovered, 3 not done
Score	76.1 of 100, leaving out posture and time
Mistakes	Case left open after the pad; wrong instrument for the needle; knot in place of the second stitch
Cosmos 3 Nano, no fine-tuning	8 of 14 grades correct
Marking every action done	9 of 14 correct, so do not judge a model on accuracy alone

A wrong object or procedure counts as not done; there are no partial grades. The full video, every grade and the score report are in execution_grading/run_1849/.

Splits

Config	Train	Validation	What it means
`default`	Protocols 1849, 1850, 1859, 1860: 78 clips + 5 augmented copies of each	Protocol 1861: 27 clips	Validation is an unseen protocol
`execution_grading`	Steps 1–3: 11 clips + 5 augmented copies of each	Steps 4–5: 3 clips	Same trainee and run, so not an unseen person

Both validation sets are small, and there is no test split. The trainer clips for the grading actions are in default train.

How it was made

Stage	Trainer clips	Trainee clips
Record	Headset, silicone pad, standard kit, at a desk	Same setup, one full run of protocol 1849
Generate	Cosmos 3 Nano, zero-shot, splits each step video into labelled actions	Cosmos 3 Nano grades each step against the trainer's reviewed labels
Review	Second by second: wrong labels fixed, missed actions added, non-actions removed	Graded windows checked at 1 s frames (2 s for two), with close-ups of the stitch
Finish	Cut, resized, audio removed, names standardised	Labels rewritten to describe only the clip; cut the same way
Augment (train)	5 copies, with mirroring where labels allow	5 copies, no mirroring

Limits

Limit	What it means
Small: one task, indoor desks, one kind of kit	Adapts a model to the format, not to other rooms, tools or procedures
Time windows checked at 0.5–1 s	A clip may start or end about 1 s off
32-frame loaders	Clips over 16 s are sampled sparsely, including 2 trainer clips over 60 s and the 76 s and 146 s trainee clips
Coarse hand poses	One of four words per hand; no joints or 3D tracking
One trainee, one run, every action attempted	Shows the format; too small to train a grader alone, and has no skipped action
Recorded by QA, no list of planned mistakes available	Grades have no independent check; some mistakes may have been planned
The headset assistant's script did not ask for the second stitch	The failed grade is real against the reference, but the trainee may have been following instructions
The two long ring-handled instruments look alike	Which is which follows the reference order; noted in `label_note`
`1849_1_a01_trainee` and `1849_2_a01_trainee` are the same 5 s	Count them once when reporting accuracy over actions
Trainee run outside the graded windows was checked more coarsely; its audio was not checked	That footage in `run_1849/` is less verified

Use

Good for: fine-tuning and evaluating structured action description; trying out grading of a practice run against written criteria; testing prompts, pipelines and output formats; step-progress and failure detection for robot learning.

Not for: clinical or medical decisions; certifying or assessing real surgical competence; use as a benchmark; identifying, tracking or profiling anyone in the videos.

Full datasets

Full datasets are graded human-hand surgery, licensed per accepted hour. Contact shar@qurexr.com.

	This sample	Full capture
Camera views	Headset only	Headset, tool close-up, room
Video	720×540, 10 fps	Headset 1024×768, up to 30 fps
Hands	Four pose words per hand	3D tracking on request
Audio	None	Time-synced narration of each step
Haptic glove	None	Time-synced force and grip data

Privacy and consent

Consent: all clips, including the trainee run, were recorded by QURE Industries, Inc., with a participant performing simulated suturing. Each participant signed a written recording release granting QURE the right to use, license and distribute the recordings, including for commercial evaluation and model training.
Not present: patients, a clinical setting, protected health information or audio.
Visible: hands, arms, wrist jewellery, a watch, a sleeve, the instruments and desk items, including a keyboard and a room camera. Protocol 1859 shows part of a laptop screen with a dashboard; only two numbers and the taskbar can be read.
Faces: none seen in a check of 4 frames from each trainer source video, or in the trainee run's review (1 s frames in graded windows, coarser elsewhere). Not every frame was checked.

License

Proprietary, under the Dataset Evaluation License (LICENSE.md). Access is gated and QURE approves each request by hand.

Internal evaluation and fine-tuning for 30 days after approval.
Models trained on it can be kept for internal use.
No redistribution. Commercial use needs a separate agreement: shar@qurexr.com.

Reference

Files

train/, validation/              default config: metadata.jsonl, videos/<id>.mp4
cosmos/                          default config in Cosmos 3 Reasoner SFT layout
execution_grading/
  train/, validation/            grading config: metadata.jsonl, videos/<id>.mp4
  cosmos/                        grading config in Cosmos 3 Reasoner SFT layout
  run_1849/                      trainee_run.mp4, judgment.json, report.json, validation.json

Sample ids are <protocol>_<step>_a<action>, with _trainee for trainee clips and _aug<n> for augmented copies.

Fields in both configs

Field	What it means
`file_name`, `id`	Video path within the split, and sample id
`protocol_id`, `step_number`, `action_id`	Where the action sits in the protocol
`duration_s`	Length in seconds
`is_augmented`, `augmentation`	Augmented copy, and its settings as JSON
`prompt`, `answer`	Instruction and target JSON

Fields in default only

Field	What it means
`source_clip`	The original clip behind a sample
`is_mirrored`	Mirrored left to right, with hands and side words swapped in the labels
`title`	Short imperative name of the action
`step_context`	One sentence on what the hands do and why, starting "The user"
`detailed_steps`	2–6 short steps
`objects_expected`	Objects used, from the 12 names below
`hand_pose_expectation`	`left` and `right`: grasping, pinching, open or idle, from the wearer's view
`success_criteria`, `failure_modes`	How the finished action looks; visible mistakes

Objects (samples): Instrument case (288), Suture pad (156), Needle holder (155), Tissue forceps (141), Suture thread (116), Suture packet (100), Suture needle (66), Plastic sleeve (49), Scalpel handle (39), Hemostatic forceps (33), Surgical scissors (33), Plastic wrap (24).

Hand poses (hands): grasping 400, pinching 393, open 186, idle 11.

Fields in execution_grading only

Field	What it means
`run_id`	The trainee run
`reference_clip`, `reference_video`	The trainer's clip for this action: id, and path from the repo root
`source_start_s`, `source_end_s`	Where the clip sits in the full run
`grade`	`done`, `recovered` or `not_done`, for filtering and balancing
`label_note`	A caveat on the label, if any

Fields inside answer:

Field	What it means
`observed`	The action was attempted
`outcome_score`	1 if the end state meets the criteria, otherwise 0
`recovered`	A wrong object or move came first, then the action was done correctly
`verdict`	What the clip shows
`deviations`	One `why` and `evidence` per mistake

Augmentation (seed 1849)

Augmentation	How often	Range
Boundary jitter	Every copy	Start and end ±0.5 s
Colour and lighting	Every copy	Brightness ±0.08, contrast ±15%, saturation ±20%, gamma ±10%, hue ±8°
Crop and scale	About 70%	Keep 85–95% of the frame
Small rotation	40%	±5°
Speed	50%	0.9–1.1×
Mirror + hand swap	50% of copies whose labels allow it (104 copies), `default` only	Left/right labels and side words swapped
Sensor noise	30%	Strength 4–10
Blur	20%	Sigma 0.6–1.4
Heavy compression	30%	H.264 CRF 30–34

To train without augmentation, keep rows where is_augmented is false.

Citation

@misc{qure2026suturing,
  author       = {{QURE Industries, Inc.}},
  title        = {Suturing Skills Action Clips},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/qure-ai/suturing-skills-action-clips}}
}

Downloads last month: 896

Total file size:

1.62 GB