livctr/oakink2-primitives

OakInk2 functional-primitive data: dataset card and provenance

Per-hand functional-primitive labels for OakInk2 [@oakink2], with real and synthetic training videos and real evaluation videos, for training and comparing action-segmentation methods (a Qwen3.5-9B LoRA, DiffAct [@diffact], FACT [@fact]). Each hand gets one of five primitives per 0.1 s step: reach, reposition, transport, stabilize, idle.

Part A is the dataset card; it ships beside the tars as README.md, with the counts filled in by python -m prims.data.export (this repository copy shows placeholders). Part B is for maintainers: how the data was made. The tars hold data only; the code, including the scorer, is this repository.

This copy: export of 2026-10-27, repository commit 3144b9f669e62cd62d011b429cfa3deecd012f5f.

Part A. Dataset card

A1. Files

File	Contents	Files inside	Bytes
labels.tar	labels/{train,val,test}.jsonl: labels v4 of every OakInk2 sequence, both hands, 0.1 s steps (A3)	4	8,151,040
train_real_full.tar	real OakInk2 training windows, native 848×480 frames (A4)	16,002	18,332,057,600
train_real_crop.tar	the same rows as two-hand 320×192 crops	16,002	4,677,181,440
train_synth_full.tar	synthetic training windows (Isaac Sim renders of OakInk2's motion), frames as rendered	16,002	9,204,633,600
train_synth_crop.tar	the same rows as 320×192 crops (the synthetic pipeline's original 16,000 crop videos)	16,002	3,857,264,640
eval_real_full.tar	real OakInk2 val and test recordings, untrimmed, 848×480, with gold (A5)	592	1,187,819,520
eval_real_crop.tar	the same recordings as 320×192 crops	592	335,534,080
SHA256SUMS	SHA-256 of README.md and of every tar

• Download with hf download <dataset> --repo-type dataset --local-dir <dir> after hf auth login (<dataset> is the id at the top of the dataset page); add file names to fetch only those, e.g. hf download <dataset> labels.tar eval_real_crop.tar --repo-type dataset --local-dir <dir>.
• Check a download with sha256sum -c SHA256SUMS; each file's size must equal its Bytes.
• Unpack with tar -xf <name>.tar: uncompressed POSIX tars, one top folder named after the tar, plain files and folders only, no links.
• Member order. Training set: the folder, index.jsonl, manifest.json, videos/, then videos/<row_id>.mp4 in index order. Eval set: the folder, then per split (val, then test) <split>/, gold.jsonl, manifest.json, videos/, then videos/<recording>.mp4 in gold order. labels.tar: labels/, train.jsonl, val.jsonl, test.jsonl, manifest.json.
• manifest.json records how a set was built (commit, recipe, inputs, counts). Its paths use the placeholders $PRIMS_OUT, $PRIMS_DATA_ROOT, $PRIMS_DETECTIONS and <repo>: provenance only.

A2. The primitives and the labels

Per hand, after the functional-primitive taxonomy [@schambra2019], from OakInk2's motion capture:

Primitive	Rule
reach	a movement still going when the hand touches an object, until the grasp is done
reposition	out of contact and moving, when the movement does not run into contact
transport	in contact after the grasp, moving or twisting the object
stabilize	in contact after the grasp, still
idle	out of contact and still

• Thresholds. Contact: the MANO hand surface [@mano] within 3 mm of an object part, released past 8 mm. Moving: the palm travels more than 2 cm over 0.5 s (in contact, also when it turns more than 10°). No StrokeRehab data or statistics set them. Full rules: docs/labeling.md.
• Labels v4 = the base rules plus (1) release: a stabilize under 2 s between a transport and letting go is transport; (2) rest stop: out of contact, a stop is idle once the palm stays under 4 cm/s for 0.3 s, except a stop with under 2 cm of travel to the touch (stays reach); a reach cut off from its contact becomes reposition.
• Two-rule merge (two_rules), on runs under 3 steps (0.3 s): (1) a short transport, reach or reposition run with the same label on both sides takes that label; (2) a short transport next to a stabilize becomes stabilize. Other short runs, and runs touching the sequence or window edge, stay. Shortest first, ties earliest, re-evaluated after each merge.
• Time. Labels cover OakInk2's annotated task span only (program_extension/endpoint/<seq>.json); outside it the arms come down from or go back to a rest pose. Step k starts k × 0.1 s into the span and is frame 3k + 1 of the 30 fps recording (frames_30fps).

Where	Labels
labels.tar and the eval gold	v4, merged once over the whole sequence (the same labels)
training rows	v4 at the steps the (tempo-warped) window shows, then merged within the window

A3. labels.tar

labels/<split>.jsonl: one line per sequence of OakInk2's official split (OakInk2-TaMF [@oakink2_tamf]), sorted by seq_id. The split is by sequence, not participant: every val and test participant also appears in train.

• train: 431 sequences, 200,098 steps;
• val: 39 sequences, 15,792 steps;
• test: 157 sequences, 68,380 steps.

{"seq_id": "A001_25281b12", "oakink2_name": "scene_02__A001/seq__25281b1248ce06ef00f4__2023-04-15-18-50-17",
 "participant": "A001", "split": "val", "fps": 10, "n_steps": 201, "frames_30fps": [1, 4, 7],
 "labels": {"left": ["idle"], "right": ["idle"]}}

Field	Notes
seq_id	<participant>_<first 8 hex digits of the sequence hash>, the id every set uses
oakink2_name	the name in OakInk2's split files
fps, n_steps, frames_30fps	10; the steps in the task span; the 30 fps frame of each step (3k + 1)
labels	left and right, n_steps labels each: v4, merged once over the whole sequence (A2); equal to the eval gold

A4. Training sets

OakInk2's official train split only. A row is one 15 s window (150 steps of 0.1 s, 10 fps) and one mp4. Real and synthetic rows follow one recipe (config/data.yaml, src/prims/data/):

Step	Real	Synthetic
Source	3 allocentric cameras, 848×480, every train sequence (A6)	renders a and b; 848×480 or 640×480, some portrait
1. Windows	13,064 random windows over (sequence, camera) in the task span	6,532 per render, over the render stretches
2. Rebalance	16,000 draws, a quarter per band of the busiest asked hand's segment count (1–2 / 3–5 / 6–10 / 11+); answer formats 20 / 20 / 30 / 30 %	8,000 draws per render
3. Quiet mix	quiet one-hand windows swapped in, alternating hands	the same
4. Labels	v4	v4, and the labels of rest-pose edits where a render replaced the motion
5. Tempo warp	20 % unchanged, else a rate drawn log-uniformly in 0.75–2×	the same
6. Merge	two-rule merge within the window	the same
7. Crop (crop sets)	YOLO26x-pose [@yolo26] on every 30 fps frame, One-Euro filter [@casiez2012oneeuro] at 30 fps in warped order, ±20 % box jitter	the same; render b with ground-truth hand assignment
8. Photometric	brightness, contrast, saturation, gamma, white balance, sharpen or blur, noise, JPEG; 20 % clean	the same

• Full and crop sets of a source share their rows (row_ids, windows, warp, labels, photometric draw); only the pixels, size, legacy_video and augment.crop_jitter differ. All augmentation is on for both sources.
• Frames. Full sets keep the native size. Crops are 320×192: the left hand's 160 px square on the left, the right hand's on the right, under a 32 px "Left hand | Right hand" title bar. Synthetic frames black out a disk over the rendered person's head; real frames are as recorded, faces included.
• Video: H.264 (yuv420p, CRF 23), 10 fps; frame k shows step k.
• Synthetic labels. In 3,279 of the 16,000 synthetic rows (20.5 %; 3,535 hands, a mean of 14 steps each, at most
  57. labels differ from v4 at the row's steps, merged within the window (what a real row with the same window would carry): these are step 4's rest-pose edits; 97 % of the changed steps go from idle to reach or reposition. The row's labels are the targets.

Set	Rows	Sequences	Frame sizes (rows)	Frames per row (rows)	Sources (rows)
train_real_full	16,000	424	848×480 (16,000)	150 (15,638), 126 (52), 87 (36), 122 (35), 101 (32), 125 (25), 112 (16), 121 (14), 78 (11), 74 (11), 106 (10), 100 (8), 110 (7), 93 (6), 136 (6), 134 (6), 143 (6), 133 (5), 82 (5), 71 (5), 132 (5), 131 (5), 79 (5), 50 (4), 89 (4), 115 (4), 108 (4), 86 (4), 142 (4), 123 (3), 81 (3), 138 (3), 114 (2), 85 (2), 145 (2), 96 (2), 146 (2), 127 (1), 147 (1), 137 (1), 98 (1), 149 (1), 109 (1), 97 (1), 84 (1)	real (16,000)
train_real_crop	16,000	424	320×192 (16,000)	150 (15,638), 126 (52), 87 (36), 122 (35), 101 (32), 125 (25), 112 (16), 121 (14), 78 (11), 74 (11), 106 (10), 100 (8), 110 (7), 93 (6), 136 (6), 134 (6), 143 (6), 133 (5), 82 (5), 71 (5), 132 (5), 131 (5), 79 (5), 50 (4), 89 (4), 115 (4), 108 (4), 86 (4), 142 (4), 123 (3), 81 (3), 138 (3), 114 (2), 85 (2), 145 (2), 96 (2), 146 (2), 127 (1), 147 (1), 137 (1), 98 (1), 149 (1), 109 (1), 97 (1), 84 (1)	real (16,000)
train_synth_full	16,000	408	640×480 (5,180), 848×480 (4,746), 480×848 (3,610), 480×640 (2,464)	150 (15,814), 126 (28), 78 (26), 112 (16), 122 (11), 110 (9), 121 (9), 125 (8), 100 (6), 132 (5), 87 (5), 93 (4), 131 (4), 134 (4), 50 (4), 79 (3), 142 (3), 82 (3), 81 (3), 106 (3), 123 (3), 133 (3), 74 (3), 136 (3), 71 (2), 114 (2), 86 (2), 117 (2), 127 (1), 101 (1), 98 (1), 113 (1), 143 (1), 97 (1), 84 (1), 115 (1), 99 (1), 145 (1), 89 (1), 146 (1)	synth_a (8,000), synth_b (8,000)
train_synth_crop	16,000	408	320×192 (16,000)	150 (15,814), 126 (28), 78 (26), 112 (16), 122 (11), 110 (9), 121 (9), 125 (8), 100 (6), 132 (5), 87 (5), 93 (4), 131 (4), 134 (4), 50 (4), 79 (3), 142 (3), 82 (3), 81 (3), 106 (3), 123 (3), 133 (3), 74 (3), 136 (3), 71 (2), 114 (2), 86 (2), 117 (2), 127 (1), 101 (1), 98 (1), 113 (1), 143 (1), 97 (1), 84 (1), 115 (1), 99 (1), 145 (1), 89 (1), 146 (1)	synth_a (8,000), synth_b (8,000)

index.jsonl, one row per video (augment.photometric shortened):

{"row_id": "synth_a_000000", "video": "videos/synth_a_000000.mp4",
 "legacy_video": "videos/A005_0ef671b3_00000_v3/front_down/00000-00150.mp4",
 "source": "synth", "render": "a", "person": "s077", "seq_id": "A005_0ef671b3", "participant": "A005",
 "camera": "front_down", "stretch": "A005_0ef671b3_00000_v3", "size": [320, 192], "fps": 10, "n_frames": 150,
 "src_steps": [0, 150], "timewarp_rate": 1.0,
 "augment": {"timewarp": true, "timewarp_rate_drawn": 1.41, "window": [0, 150],
             "crop_jitter": {"left": [0.003, -0.112], "right": [0.086, 0.073]},
             "photometric": {"clean": true, "brightness": 1.0, "jpeg_quality": null, "seed": 237542308}},
 "vlm_format": "right_left", "labels": {"left": ["reposition"], "right": ["reach"]}}

Field	Notes
row_id, video	real_000123, synth_a_000123, synth_b_000123; the mp4 is videos/<row_id>.mp4
size, n_frames	[width, height]; 150, fewer when the source is under 15 s (a synthetic stretch, or a whole short real sequence, window [0, n_steps))
labels	left and right, n_frames labels each
src_steps, timewarp_rate	the source steps [first, end) shown (synthetic: of the stretch; sequence step = stretch start + stretch step); the tempo shown (> 1 faster, 1.0 unchanged)
augment	the draws: timewarp_rate_drawn (before fitting to the source), window (the source window, before the warp), crop_jitter (per hand [dx, dy] in box sides; null in full sets), photometric (clean: unchanged)
vlm_format	left, right, left_right, right_left (20 / 20 / 30 / 30 %): the hand(s) the VLM is asked about. Every row labels both hands (methods/README.md, "Answer formats")
source, render, person, seq_id, participant, camera, stretch	provenance: real or synth; render a or b and the rendered person (not an OakInk2 participant); the OakInk2 sequence; real camera allocentric_top, _right or _left, or the render view; stretch <seq_id>_<start step, 5 digits>_v3 (a) or _v3b (b)
legacy_video	synthetic crop rows: the original build's path; ignore it

Tempo warp. Output step k shows the source step nearest to the window's centre + (k − (n − 1)/2) × timewarp_rate (n = n_frames), the run shifted to lie inside the source. prims.data.timewarp.source_steps(window[0], window[1] - window[0], <source length>, timewarp_rate_drawn) returns (steps, rate). <source length> is n_steps (labels.tar) for real rows; for synthetic rows the stretch length is not shipped, so use src_steps.

A5. Eval sets

OakInk2's official val (39 sequences) and test (157) splits, each seen by its 3 allocentric cameras. A recording is one (sequence, camera), id <seq_id>__<camera>: one untrimmed 10 fps mp4 of the whole task span.

	val recordings	val sequences	val steps	test recordings	test sequences	test steps
eval_real_full	117	39	47,376	471	157	205,140
eval_real_crop	117	39	47,376	471	157	205,140

• Frame k is the camera image at 30 fps frame 3k + 1 (exceptions: A6). No augmentation.
• eval_real_full is 848×480; eval_real_crop is 320×192 crops by the training crop rule, smoothed over the whole recording, without jitter. Both have the same recordings and gold: v4 of both hands, merged once over the whole sequence (A2).

<split>/gold.jsonl, one row per recording; sequences sorted, cameras in the order top, right, left. n_steps equals the video's frame count and the length of frames_30fps and of each labels list:

{"recording": "A001_25281b12__allocentric_top", "video": "videos/A001_25281b12__allocentric_top.mp4", "split": "val",
 "seq_id": "A001_25281b12", "participant": "A001", "camera": "allocentric_top", "fps": 10, "n_steps": 201,
 "frames_30fps": [1, 4, 7], "labels": {"left": ["idle"], "right": ["idle"]}}

A6. Missing camera images in six sequences: the nearest-image rule

Six sequences lost camera images, at the same moments on all three cameras: O001_77461f70, O003_5a9d8e3c, O003_b4972805, O003_c3dbd1f0, O003_eb87745d (train) and O003_2a060a99 (test). Each gap is one dropped camera interval (83–217 ms between images). Every other sequence has an image within one 30 fps frame of every label frame.

The rule: every frame and every 30 fps pose row uses the camera's nearest image within 0.1 s; anything farther fails the build (the worst offset is 91.7 ms). Every recording is kept whole.

• Eval: in O003_2a060a99's 10 fps videos, steps 146 and 232 are 41.7 ms off.
• Training: 77 of the 13,064 pool windows touch a gap. In the drawn rows, one to seven steps show an image up to 58 ms off: 66 rows of train_real_full, and 74 of train_real_crop, whose 8 more rows have only a crop pose taken from a neighbouring image. Crop boxes there come from poses on the nearest images, up to 92 ms off.
• Each set's manifest.json lists the affected recordings, steps and offsets (eval: params.nearest_image_recordings). Test: O003_2a060a99__allocentric_left, O003_2a060a99__allocentric_right, O003_2a060a99__allocentric_top. Val: none.

A7. Scoring with prims-eval

Every method is scored by prims_eval (eval/ in this repository; numpy only):

git clone <this repository> && cd <repository>/eval && uv sync
uv run prims-eval windows --gold eval_real_full/test/gold.jsonl --out windows.jsonl
# predict every listed window from its own frames (start_step to end_step - 1 of the recording's mp4), both hands
uv run prims-eval score --gold eval_real_full/test/gold.jsonl --pred predictions.jsonl --out scores/

predictions.jsonl has one line per window per hand, with one label (or null, scored as wrong) per step: {"recording": "A001_25281b12__allocentric_top", "hand": "left", "start_step": 120, "labels": ["idle", "reach", null]}. Select models on val, report test once; test full-frame models on eval_real_full and crop models on eval_real_crop. Windows, stitching and metrics (edit score, AER, RCE [@vlms4rehab], accuracy, macro-F1): eval/README.md. The method contract: methods/README.md.

A8. Licences and use

The data, and the third-party parts the repository uses:

What	Licence	What it means here
OakInk2 [@oakink2]: every set here (real frames; the motion the renders replay)	CC BY-SA 4.0	credit OakInk2 (Zhan et al., CVPR 2024); share derivatives under the same licence
MANO [@mano]: the labels come from MANO meshes; the renders show MANO hands	non-commercial research only	No MANO model files are included. Use train_synth_* for non-commercial scientific research only, as MANO's licence requires, and cite MANO
Render people, therapy-worlds [@therapyworlds]	CC0 and CC BY MakeHuman assets	credit lines in therapy-worlds' subjects/v2/manifest.json
Poly Haven textures and HDRIs [@polyhaven2026]	CC0
Monk Skin Tone Scale [@monk2022]	CC BY 4.0	skin tones in config/synth.yaml
Isaac Sim 6.1 [@isaacsim2026]	Apache-2.0; Kit SDK under the NVIDIA Omniverse License Agreement	the renderer; not shipped
YOLO26x-pose [@yolo26]	AGPL-3.0	crop boxes; no weights shipped
Qwen3.5-9B [@qwen35]	Apache-2.0	the VLM's base model
DiffAct [@diffact], FACT [@fact]	MIT	upstream code, pinned, not vendored
StrokeRehab [@strokerehab], FP90 [@vlms4rehab]	not shipped	used only where its data may be used (B4)

• Personal data. The real frames show OakInk2 participants' upper bodies; faces are visible in some views (clearly in allocentric_left, partly in allocentric_top). No face crops, no re-identification; keep review pages private.
• StrokeRehab. Never view StrokeRehab video or frames, in any form (no frame dumps, crops, contact sheets or video pages); read only labels, decodes and scores. FP90 is numeric only.

A9. Citations

Cite OakInk2 [@oakink2], its official split [@oakink2_tamf], MANO [@mano], the primitive taxonomy [@schambra2019], and vlms4rehab [@vlms4rehab] for the metrics. Every key is in docs/references.bib:

@inproceedings{oakink2,
  title     = {{OakInk2}: A Dataset of Bimanual Hands-Object Manipulation in Complex Task Completion},
  author    = {Zhan, Xinyu and Yang, Lixin and Zhao, Yifei and Mao, Kangrui and Xu, Hanlin and Lin, Zenan and Li, Kailin and Lu, Cewu},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2024}
}
@article{mano,
  title   = {Embodied Hands: Modeling and Capturing Hands and Bodies Together},
  author  = {Romero, Javier and Tzionas, Dimitrios and Black, Michael J.},
  journal = {ACM Transactions on Graphics (Proc. SIGGRAPH Asia)},
  volume  = {36}, number = {6}, year = {2017}, doi = {10.1145/3130800.3130883}
}
@article{schambra2019,
  title   = {A Taxonomy of Functional Upper Extremity Motion},
  author  = {Schambra, Heidi M. and Parnandi, Avinash and Pandit, Natasha G. and Uddin, Jasim and Wirtanen, Audre and Nilsen, Dawn M.},
  journal = {Frontiers in Neurology}, volume = {10}, pages = {857}, year = {2019}, doi = {10.3389/fneur.2019.00857}
}
@article{vlms4rehab,
  title   = {Vision-language models for human motion understanding: Lessons from stroke rehabilitation},
  author  = {Li, Victor and Kamalakannan, Naveenraj and Parnandi, Avinash and Schambra, Heidi and Fernandez-Granda, Carlos},
  journal = {PLOS Digital Health}, volume = {5}, number = {7}, pages = {1--19}, year = {2026},
  doi     = {10.1371/journal.pdig.0001506}
}

A10. Checks anyone can run on the tars

These hold for every export; a failure means the export is broken.

1. sha256sum -c SHA256SUMS passes; each tar's size equals A1's Bytes.
2. Each tar has exactly A1's member order and "Files inside" regular files, all under its own top folder.
3. Training sets: A4's row count; unique row_id; every row's video is a member and no other mp4 is; labels.left and labels.right have n_frames entries from the five primitives; each mp4 decodes to n_frames frames of size. A source's full and crop sets have the same row_ids and, per row, the same labels, src_steps, timewarp_rate, vlm_format and augment.photometric.
4. Eval sets: A5's recording counts; frames_30fps[k] == 3k + 1; n_steps equals the length of frames_30fps, of each labels list and the mp4's frame count; full and crop gold rows are identical; videos are 848×480 (full) or 320×192 (crop).
5. Labels: A3's sequence counts; each val or test sequence's labels equals the labels of its three recordings in both eval sets' gold.

Part B. Provenance

How the data in Part A, and the data vlm_synth_crop (the VLM trained on train_synth_crop; run ID xI20261017_v3x2_lab4, checkpoint-2000) used, were made. Commands: README.md; details: the modules' docstrings.

• Paths use the site variables of config/site.env: $PRIMS_DATA_ROOT (data, default data/), $PRIMS_OUT (outputs, default out/; sets in $PRIMS_OUT/sets/<set>/), $PRIMS_DETECTIONS (pose tracks, default $PRIMS_OUT/detections). None of it is in git.
• The original build. Bare script names below (prims-sft build, rebalance_rate.py, quiet_mix.py, bundle_ctx_dataset.py, concat_sets.py, keep_views.py, crop_ctx_dataset.py, sr_c30_crops.py, sr_metrics.py) are in git tag pre-cleanup-2026-10-26; their outputs stay in $PRIMS_OUT/sft/data/. The ports (src/prims/data/, eval/, methods/vlm/) were each checked against those outputs.

B1. OakInk2 [@oakink2]

From the Hugging Face dataset kelvin34501/OakInk-v2 (not gated). Video is one tar per sequence: 627 tars, 2.09 TB, 0.6–20 GB each. Local layout, $PRIMS_DATA_ROOT/oakink2/ (all 627 sequences):

Path	What (made by)
anno_preview/<seq>.pkl	torch pickles: MANO, objects, SMPL-X, cameras (downloaded; opened only by scripts/convert_oakink2.py's allow-list unpickler)
converted/<seq>.npz	the labeler's input, 120 fps, y-up (scripts/convert_oakink2.py)
converted_body/<seq>.npz	the SMPL-X upper-body fit, for the renderer (scripts/extract_oakink2_body.py)
program/, program_extension/endpoint/	per-hand primitive-task annotations; the annotated task span (downloaded)
object_repair/align_ds/<id>/model.obj	object meshes; fallback object_affordance/affordance_part/<id>/*.obj (downloaded)
splits/tamf/	the official split (below)
video/<full sequence name>/	per camera <serial>.mp4 (30 fps, H.264 CRF 18) and <serial>.json; cameras.npz; manifest.json last (scripts/pack_oakink2_video.py)

• Fetch: scripts/cluster/submit_all.sh runs convert, fetch, pack and prims label per sequence, skipping finished work. Tars are untrusted: only <seq>/<digits>/<digits>.png members are extracted (filter="data"), nothing is executed, and the tars are deleted afterwards.
• Annotations (no tracked script downloads them): uv run --with huggingface_hub python -c "from huggingface_hub import snapshot_download as d; d('kelvin34501/OakInk-v2', repo_type='dataset', allow_patterns=['anno_preview/*'], local_dir='$PRIMS_DATA_ROOT/oakink2')" (likewise program*/, object_*/).
• Task span: program_extension/endpoint/<seq>.json [seq_beg, seq_end) in mocap frames; outside it (3–8 s before, 4–13 s after, on the first four sequences checked) the arms move from or to a rest pose, unannotated.
• Dropped frames (A6): the original 10 fps builds refused O003_b4972805, O003_c3dbd1f0 and O003_eb87745d (real video of 428 of the 431 train sequences); Part A's sets keep all six sequences (config/data.yaml real.max_frame_offset_s; poses in $PRIMS_DETECTIONS/oakink2_allo30 and oakink2_allo30_nearest). The renders use all 431.

Splits [@oakink2_tamf]: 431 train (333.5 min), 39 val (26.3 min), 157 test (114.0 min). Split files name sequences with /, data folders with ++. scripts/synth_split.py --parts train writes the renderer's split.

mkdir -p $PRIMS_DATA_ROOT/oakink2/splits/tamf && cd $PRIMS_DATA_ROOT/oakink2/splits/tamf
for f in train val test all; do
  curl -fsSLO "https://raw.githubusercontent.com/oakink/OakInk2-TaMF/ce018173e0981fec43d1f53328e06808cf00198c/asset/split/$f.txt"
done
git hash-object train.txt val.txt test.txt all.txt   # bbff714e35c6..., b620d8eb45ca..., 3131fc99323e..., c523f7626d6b...

B2. Synthetic renders (the synthetic pipeline)

Two Isaac Sim renders of the 431 train sequences in the same 573 stretches (config/synth.yaml, renders; details in EXPERIMENT_PLAN.md, "Rendering details"):

Render	Clips	Seed	Stretches rendered	Detections (scripts/detect_windows.py --kind synth)
a	$PRIMS_OUT/synth/clips_v3full	3	569	$PRIMS_DETECTIONS/synth_v3full
b	$PRIMS_OUT/synth/clips_v3full_b	4	566	$PRIMS_DETECTIONS/synth_v3full_b

• Inputs: converted/, converted_body/, program_extension/endpoint/ and object meshes; people, bystanders and their motions from $PRIMS_TW_DATA_ROOT [@therapyworlds] (subjects/v2, clinicians/v1, assets/v3/motions; scripts/cluster/tw_people.py, tw_motions.py).
• Renderer: Isaac Sim 6.1.0 ($ISAAC_SIF, scripts/cluster/isaac_exec.sh) on any GPU with RT cores (L40S used).
• Labels: base rules $PRIMS_OUT/labels/oakink2_all (prims label); unmerged v4 $PRIMS_OUT/labels/oakink2_v4 (scripts/relabel_steps.py).
• Original training set $PRIMS_OUT/sft/data/synth_v3x2lab4_noctx_none (16,000 rows; commit 8aa0e41, dirty): per render prims-sft build → rebalance_rate.py → quiet_mix.py → bundle_ctx_dataset.py --relabel $PRIMS_OUT/labels/oakink2_v4 (synth_v3full_lab4_noctx_none, synth_v3fullb_lab4_noctx_none), then concat_sets.py. train_synth_crop hard-links the same 16,000 mp4s (python -m prims.data link-legacy; recipe key synth.renders.<render>.legacy_set).

B3. The earlier OakInk2 eval set

$PRIMS_OUT/sft/data/oakink2_tamf_allo_crop26oe_noctx_none: val 39 sequences and 846 one-hand answers, test 157 and 3,642, on the three allocentric cameras. A5 replaces it; it stays the reference for vlm_synth_crop's numbers before the re-decode on eval_real_crop. Chain: prims-sft build --source oakink2_video --train-from oakink2_a100 → keep_views.py → crop_ctx_dataset.py --highlight none --ctx-key handcrops_v2_noctx, on detections $PRIMS_DETECTIONS/oakink2_allo (scripts/cluster/handdet_detect26x.sbatch, scripts/detect_plan.py plan allo_rate4k_train_valtest.json).

B4. StrokeRehab, FP90 and StrokeRehab 2022 eval sets (frozen; not exported)

Only where the StrokeRehab data may be used. Never view StrokeRehab video or frames, in any form; FP90 is numeric only. Clips and pose tracks stay under that site's $PRIMS_OUT and $PRIMS_DETECTIONS, unchanged, so vlm_synth_crop's numbers stay exact.

Set ($PRIMS_OUT/sft/data/)	What	Videos	Scored windows	Context-only windows
sr_c30oe_r10	StrokeRehab [@strokerehab], all 71 participants, one camera per recording	2,861	11,676	489 (307 units)
fp90_c30oe_r10	vlms4rehab FP90 [@vlms4rehab]: 90 videos of 10 subjects (5 stroke, 5 control)	90	408	0
sr2022_c30oe_r10	StrokeRehab 2022 test (rehabbench in_strokerehab2022_test): 8 stroke participants, 260 recordings of reps 1–5, both cameras	514	2,157	22 (20 units)

• Gold for prims_eval: python -m prims.data.strokerehab (see its docstring) writes $PRIMS_OUT/sets/eval_{strokerehab,fp90,sr2022}_crop over these clips; score with prims-eval score --protocol strokerehab|fp90|sr2022.
• Windows: rehabbench's [@rehabbench] StrokeRehab manifest (raw .mkv, 60 fps, two cameras), 15 s every 12 s at 10 fps, on the chosen camera; the tested arm only; unlabelled spans are ignore, and all-ignored windows are dropped. Context-only windows are never scored.
• Built by scripts/cluster/sr_c30_crops_full.sbatch (tag): sr_c30_crops.py plan, run, merge --highlight none --allow-partial; questions copied from the original crop sets (sr_rep1_raw_crop26x, sr_rest_raw_crop26x, vlms4rehab_fp90_crop26x). Plans and clips in $PRIMS_OUT/sr_c30oe_cache/.
• Crops, as in src/prims/data/cropclip.py: YOLO26x-pose at 30 fps ($PRIMS_DETECTIONS/sr_sel30), the patient tracked from the frame-0 pick, One-Euro (0.1 Hz, β 6) over the whole video; step m is pose row 3m; 320×192, 10 fps, CRF 23, last frame repeated past the end; no markers or tint.
• Camera and patient (view run r10, $PRIMS_OUT/view_exp/runs/r10; code in the tag). On frame 0, YOLO26x-pose finds people and Qwen3.5-9B [@qwen35] picks the patient and matches them across cameras; it also says which cameras are Aerial, and SAM 3D Body [@sam3dbody] gives alpha, the camera's angle around the patient (0° in front, positive toward the tested hand). Rule: the only Aerial camera; else the larger alpha (shelf: alpha nearest 25°); for combing, deodorant, drinking, feeding, glasses and RTT the level camera wins when its alpha > 0; both Aerial: camera 1.
• Agreement and pinned list. One annotator's camera and patient labels for all 2,861 recordings (patient_camera_labels_20261015.*; 7 excluded with a note) match r10 on all 2,854 others; nothing read them at run time. rehabbench 90b45fe pins them (fp_recordings.csv, excluded=False, chosen_camera). On the 2,854, vlm_synth_crop scores ARE 11.26 ± 7.98 and ES 57.34 ± 9.96.
• Original scoring (scripts/report/sr_metrics.py, tag), reproduced by prims-eval score --protocol strokerehab (eval/README.md). Its headline averages both cameras per recording; a policy-camera variant uses rehabbench's view_preference_order with fp_views.csv (humans checked 382 videos; the model labels agreed on 79 %). The ARE CI is a 1,000-resample participant bootstrap. format_ok reads 0.000 on these sets (they store the "arm" answer header); labels are unaffected.

B5. The export

python -m prims.data.export build | verify | upload (src/prims/data/export.py; scripts/cluster/export.sbatch; --help for flags). build and upload are dry runs unless --execute.

• build checks each set, writes one uncompressed tar per set to $PRIMS_OUT/export/ (linked videos stored as files), checks each tar against its plan and keeps its SHA-256; once all seven tars match their sets it writes README.md (this file, counts filled in) and SHA256SUMS. verify [--hash] re-checks them.
• Shipped manifests have the resolved site paths and the repository replaced by placeholders (plus $PRIMS_EXPORT_SCRUB); the build fails if a word of $PRIMS_EXPORT_DENY is left in any shipped text, this card included.
• The dataset is published as a public Hugging Face dataset: hf upload <dataset> $PRIMS_OUT/export . --repo-type dataset --include "*.tar" README.md SHA256SUMS. For any other destination, upload --remote REMOTE:PATH runs rclone copy ($RCLONE or PATH) of the same files, then compares each remote file's size and SHA-256.