README.md

metadata

license: other
license_name: per-source-terms
pretty_name: Pretext-270K
language:
  - en
size_categories:
  - 100K<n<1M
task_categories:
  - image-to-image
  - any-to-any
configs:
  - config_name: jigsaw
    data_files:
      - split: train
        path: preview/jigsaw.parquet
  - config_name: orthogonal_transform
    data_files:
      - split: train
        path: preview/orthogonal_transform.parquet
  - config_name: deformation_recovery
    data_files:
      - split: train
        path: preview/deformation_recovery.parquet
  - config_name: super_resolution
    data_files:
      - split: train
        path: preview/super_resolution.parquet
  - config_name: colorization
    data_files:
      - split: train
        path: preview/colorization.parquet
  - config_name: denoising
    data_files:
      - split: train
        path: preview/denoising.parquet
  - config_name: masked_inpainting
    data_files:
      - split: train
        path: preview/masked_inpainting.parquet

Pretext-270K

Pretext-270K is the first-stage training data of ThinkMorph-SSL-Base. It holds 270,896 training traces and 1,600 validation traces. Each trace starts from a corrupted input image. A description follows, then the restored image, then a further description. There is no question and no answer.

The traces are grouped into seven tasks, one per corruption that the model undoes.

Tasks

Task	Input image	Restored image	Train rows	Val rows	Of these, text only
jigsaw	shuffled tiles: one image cut into tiles, or separate tile images	reassembled image	52,975	571	10,266
orthogonal_transform	rotated or flipped image	upright image	36,763	183	—
deformation_recovery	geometrically warped image	unwarped image	36,672	189	—
super_resolution	low-resolution image	high-resolution image	36,409	179	—
colorization	grayscale image	colored image	36,104	168	—
denoising	noisy image	clean image	36,023	155	—
masked_inpainting	image with masked regions	completed image	35,950	155	—
Total			270,896	1,600	10,266

Stage 1 trained on the train rows only. Choose a task in the Dataset Viewer to see examples.

Files

Path	Content
shards/annotations.tar	MANIFEST.json and every task file: all rows, no images
shards/data-*.tar	the 533,366 images (141 GB) in 133 tar shards
shards/SHARDS.json	size and SHA256 of every shard
text_only/	10,266 rows whose images (Matterport3D) are not distributed here; see Rows without images
preview/	the Dataset Viewer examples, one file per task

Download

hf download ThinkMorph/Pretext-270K --repo-type dataset --include "shards/*" --local-dir Pretext-270K
mkdir -p data
for f in Pretext-270K/shards/*.tar; do tar -xf "$f" -C data; done

• shards/SHARDS.json gives each shard's SHA256, and MANIFEST.json each extracted file's.
• To read the rows without the images, download and extract shards/annotations.tar only.

Layout and row format

After extracting the shards:

data/
  MANIFEST.json
  images/<sha256>.<ext>          shared by all tasks
  colorization/train.jsonl
  colorization/val.jsonl
  ...                            one folder per task

{"id": "<sha256>", "task": "colorization", "images": ["../images/<sha256>.jpg"],
 "reasoning_thought_0": "...", "reasoning_image_0": "../images/<sha256>.jpg",
 "reasoning_thought_1": "..."}

• images is the corrupted input. reasoning_image_0 is the restored image.
• Image paths are relative to the row's file. id is a stable public id.

Rows without images

text_only/ holds the rows of the jigsaw task whose images come from Matterport3D. It has the same task folders as the main release. The terms of their image sources do not allow us to redistribute the images.

• text_only/images.jsonl names every image. For each one it gives the upstream identifiers, how the image was derived, and the size and MD5 of the file used in training.
• Put the images you obtain in text_only/images/; the rows refer to them as ../images/<name>.

To obtain the source data:

• Matterport3D: sign the Matterport3D Terms of Use. The form is at https://niessner.github.io/Matterport/; send it to matterport3d@googlegroups.com.

License

The images come from several public datasets, and each keeps its own terms. There is no blanket license for the images. A list of the image sources per task will be added with the paper.