Pretext-270K

Pretext-270K is the Stage 1 data of PRIMER, the recipe used to train ThinkMorph-SSL. It has 270,896 instruction-free text-image-text traces. Each trace starts from a corrupted input image. A textual description follows, then the clean target image, then a further description.

• Images: the target images come from the original source images.
• Text: a teacher model (GPT-5) writes the text segments. It first captions the clean image, then sees the caption and both images. The teacher therefore knows the target, which the student does not see in its input.

Status: private staging copy. Redistribution terms differ by source and are not all cleared; see Source terms. Do not make this repository public until the owner signs off on every source.

Subsets

Subset	Train rows	Val rows	Images	GB	Shards	Loader check	Terms status	source field (top)
mixed_100k	99,600	400	182,287	36.88	35	PASS	pending	ytb_roomtour 27,828, hypersim 24,852, deepvision-visual_logic-26k 8,482, deepvision-math-77k 5,240, +160 more
jigsaw_fragments_18k	17,951	400	98,249	—	—	text only (no image bytes)	text-only	matterport_3d_pairs_merged 10,266, realsee3d_pairs_merged 7,405, 360+x (field reads stage0) 453, 360x_pairs_merged 227
sa1b_55k	55,042	400	110,800	46.21	44	PASS	pending	SA-1B 55,442
general_99k	98,303	400	197,054	56.47	53	PASS	pending	PixMo-Cap 98,703
Total	270,896	1,600	588,390	139.57	132

Stage 1 trains on the train rows only (include_val_in_training is off for Stage 1). That is 270,896 rows, which is the paper's count. The 1,600 val rows, 400 per subset, are published as a separate split and were not used for training.

Counts by transformation family (train rows)

Family (task field)	Rows	Paper
Jigsaw restoration (jigsaw 35,024 + jigsaw_fragments 17,951)	52,975	52,975
Colorization	36,104	36,104
Super-resolution	36,409	36,409
Masked inpainting	35,950	35,950
Rotation/flip recovery (orthogonal_transform)	36,763	36,763
Denoising	36,023	36,023
Deformation recovery	36,672	36,672
Total	270,896	270,896

Counts by source (train rows) and the 443-row difference

The paper's source table lists eight sources. It adds up to 270,453, which is 443 fewer than the family total of 270,896. The data explains the gap:

Source	Rows in data	Paper source table
PixMo-Cap (general_99k)	98,303	98,303
SA-1B (sa1b_55k)	55,042	55,042
mixed_100k: VSI-590K + LLaVA-OneVision-1.5 + DeepVision-103K	99,600	58,546 + 27,379 + 13,675 = 99,600
Matterport3D (jigsaw_fragments_18k)	10,035	10,035
Realsee3D (jigsaw_fragments_18k)	7,251	7,251
360+x (jigsaw_fragments_18k, source = 360x_pairs_merged)	222	222
360+x (jigsaw_fragments_18k, source = stage0)	443	not listed
Total	270,896	270,453

In mixed_100k the source field is finer than the paper's three pools (ytb_roomtour, hypersim, arkitscenes, scannet, ... for VSI-590K; many LLaVA-OneVision-1.5 component names). The DeepVision part can be counted directly: deepvision-visual_logic-26k 8,455 + deepvision-math-77k 5,220 = 13,675.

The 443 rows are 360+x images: their image names and original_image metadata point to 360x_pairs_merged/stage0/.... Only their source field says stage0, which is why the source table missed them. The data is left exactly as it was trained on.

Download and restore

Each subset folder holds SHARDS.json and deterministic tar shards. MANIFEST.json inside the shards lists every file with its SHA256. Use the PRIMER code release:

python scripts/download_release.py <subset>/SHARDS.json \
  --base-url https://huggingface.co/datasets/ThinkMorph/<REPO>/resolve/<COMMIT>/<subset> --output shards/<subset>
python scripts/shard_data_release.py restore shards/<subset>/SHARDS.json shards/<subset> DATA_ROOT/<corpus>/<subset>

download_release.py checks each shard's size and SHA256. restore checks every shard before extracting it, then checks every file against MANIFEST.json. A restored subset contains train.jsonl, val.jsonl (when there is one), images/ and MANIFEST.json, which is the layout the BAGEL loader reads.

Row format

{"id": "<sha256>", "question": "...", "options": "...", "images": ["images/<sha256>.jpg"],
 "reasoning_thought_0": "...", "reasoning_image_0": "images/<sha256>.jpg",
 "reasoning_thought_1": "...", "answer": "..."}

• The supervision text (question, options, answer/gt_answer, every reasoning_thought_N) and the image bytes are exactly as in the training copy.
• Only two things change: image paths become content-hashed relative paths, and rows get public ids.
• Source metadata fields are removed, because they contain internal storage paths.
• No rows are filtered, no answers are rewritten and no images are recompressed. The only exceptions are listed under "Left out".

Subsets without image bytes

These subsets are published as text only: jigsaw_fragments_18k. Their image sources do not allow the images to be redistributed in this dataset. Each of these folders holds:

• train.jsonl / val.jsonl: the rows, normalized like the other subsets (same public ids and text). The image fields keep the original file names, so every image a row uses is named.
• images.jsonl: one record per referenced image, with
  • upstream_dataset and its terms;
  • upstream_ids (Matterport3D panorama id, Realsee3D scene and viewpoint, 360+x video id and frame, Ego-Exo4D take name, camera and frame rank);
  • role (for example left, right, merged, tile_03, clean, ego, exo);
  • a derivation sentence;
  • the byte size and MD5 of the exact file used in training.
• provenance.jsonl: one record per row, with the row's source, task and metadata. Storage locations are removed. What remains includes the reconstruction parameters: yaw angles of the paired views, jigsaw grid size and tile permutation, original image id and size, take name and frame rank.
• MANIFEST.json, FILES.json: file list with SHA256. FILES.json is the index format read by scripts/download_release.py.

To rebuild the images, obtain the upstream dataset under its own terms. Regenerate each file from the identifiers and parameters above, and check it against the recorded size and MD5. The generation code for the paired views and jigsaw tiles is not part of this release, so a rebuilt image may differ in bytes even when it shows the same content. Ego-Exo4D names give the take and frame rank but not the exo camera or the timestamp.

Left out

Nothing. Every row of the four training files and the four val files is included.

Source terms

Subset	Status	Sources and declared terms
mixed_100k	pending	Hypersim (CC BY-SA 3.0); YouTube room-tour frames (no redistribution grant found); DeepVision-103K (MIT); Wikipedia images (per-file licenses); CLEVR (CC BY 4.0); robotics / chart and other mixed pools (terms not recorded)
jigsaw_fragments_18k	text-only	Matterport3D EULA for academic use (non-commercial; click-wrap for substantial portions); Realsee3D (Zenodo record lists CC BY 4.0; coverage of the images unconfirmed); 360+x (CC BY-NC-SA 4.0 on the project page; CC BY 4.0 on the eData record)
sa1b_55k	pending	SA-1B (Meta SA-1B dataset research terms; distinct from SAM's Apache-2.0 code license)
general_99k	pending	PixMo-Cap (ODC-BY-1.0 for the annotations; images are third-party web images distributed by URL)

pending means the terms have not been checked against the exact upstream versions in use. restricted means the source's terms do not allow public redistribution of its images. A permissive code license (for example SAM's Apache-2.0) does not license the images, and no blanket license is claimed for the images here.

Provenance

• Training copy. Training read an internal storage copy that is no longer reachable.
• This release. It is built from the team's archived copy of the same training folders.
• What was checked. Train row counts, task-family counts and source counts match the paper's records exactly.
• What could not be checked. Byte identity with the training copy. The manifests therefore record training_annotation_identity_verified: false.