Datasets:

ThinkMorph
/

Instruct-180K

license: other
license_name: per-source-terms
pretty_name: Instruct-180K
language:
  - en
size_categories:
  - 100K<n<1M
task_categories:
  - visual-question-answering
  - any-to-any
configs:
  - config_name: birds_eye_view_rotation
    default: true
    data_files:
      - split: train
        path: preview/birds_eye_view_rotation.parquet
  - config_name: panoramic_completion
    data_files:
      - split: train
        path: preview/panoramic_completion.parquet
  - config_name: jigsaw_multi_frame
    data_files:
      - split: train
        path: preview/jigsaw_multi_frame.parquet
  - config_name: jigsaw_single_image
    data_files:
      - split: train
        path: preview/jigsaw_single_image.parquet
  - config_name: visual_search_bbox
    data_files:
      - split: train
        path: preview/bbox.parquet
  - config_name: object_rotation
    data_files:
      - split: train
        path: preview/object_rotation.parquet
  - config_name: object_orientation
    data_files:
      - split: train
        path: preview/object_orientation.parquet
  - config_name: crop
    data_files:
      - split: train
        path: preview/crop.parquet
  - config_name: depth_map
    data_files:
      - split: train
        path: preview/depth_map.parquet
  - config_name: marked_position_view
    data_files:
      - split: train
        path: preview/marked_position_view.parquet
  - config_name: object_segmentation
    data_files:
      - split: train
        path: preview/object_segmentation.parquet
  - config_name: object_viewpoint_render
    data_files:
      - split: train
        path: preview/object_viewpoint_render.parquet
  - config_name: marked_depth_map
    data_files:
      - split: train
        path: preview/marked_depth_map.parquet
  - config_name: object_viewpoint_render_oriented
    data_files:
      - split: train
        path: preview/object_viewpoint_render_oriented.parquet
  - config_name: object_editing
    data_files:
      - split: train
        path: preview/object_editing.parquet
  - config_name: birds_eye_view
    data_files:
      - split: train
        path: preview/birds_eye_view.parquet

Instruct-180K

Instruct-180K is the instruction-tuning data of ThinkMorph-SSL-Instruct, trained from ThinkMorph-SSL-Base. It holds 180,490 interleaved reasoning traces. Each trace has a question, one or more input images, reasoning text with intermediate images, and a final answer.

The traces are grouped into 17 tasks by what their intermediate images do. The levels follow the paper: basic visual perception, static spatial understanding, and mental world modeling (object manipulation and viewpoint synthesis).

Tasks

Level	Task	Intermediate images	Rows with images	Text-only rows
L1 Basic visual perception	`bbox`	a box drawn around the target	14,686	—
	`crop`	a zoomed-in crop of the target	7,479	—
	`jigsaw_single_image`	the pieces of one image, reassembled	19,935	—
	`jigsaw_multi_frame`	separate image pieces, reassembled into one image	20,000	—
L2 Static spatial understanding	`object_segmentation`	the objects to count, masked out	5,569	—
	`depth_map`	a depth map	6,849	—
	`marked_depth_map`	the query points marked on the image, then a depth map	2,726	—
L3 Object manipulation	`object_rotation`	the object after each camera rotation	10,811	—
	`object_editing`	the scene after the edits	1,000	—
L3 Viewpoint synthesis	`birds_eye_view`	a bird's-eye view of the scene	284	—
	`birds_eye_view_rotation`	a bird's-eye view, then re-oriented views	8,371	—
	`panoramic_completion`	the views completed into a panorama	24,247	33,631
	`object_viewpoint_render`	object masks, then the view from one object toward a point	4,639	—
	`object_viewpoint_render_oriented`	object masks, the object's 3D axes, then the view along its facing direction	1,162	—
	`object_orientation`	an object mask, then its 3D axes	8,063	—
	`marked_position_view`	the view from a marked spot after turning (rendered rooms)	6,512	—
	`exo_to_ego`	the person's own view, from an outside view	—	4,526
Total			142,333	38,157

All rows are training rows. Choose a task in the Dataset Viewer to see examples; there bbox is listed as visual_search_bbox.

Files

Path	Content
`shards/annotations.tar`	`MANIFEST.json` and every task file: all rows, no images
`shards/data-*.tar`	the 286,889 images (157 GB) in 148 tar shards
`shards/SHARDS.json`	size and SHA256 of every shard
`text_only/`	38,157 rows whose images (Matterport3D, Ego-Exo4D) are not distributed here; see Rows without images
`preview/`	the Dataset Viewer examples, one file per task

Download

hf download ThinkMorph/Instruct-180K --repo-type dataset --include "shards/*" --local-dir Instruct-180K
mkdir -p data
for f in Instruct-180K/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
  bbox/train.jsonl
  crop/train.jsonl
  ...                            one folder per task

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

options appears only in tasks that keep the choices outside the question.
Image paths are relative to the row's file. id is a stable public id.

Reading the data

A trace reads in this order: the input images and the question, then reasoning_thought_0, reasoning_image_0, reasoning_thought_1, and so on, then the answer.

import json
from pathlib import Path
from PIL import Image

folder = Path("data/bbox")                      # one task folder of the extracted release
with open(folder / "train.jsonl") as f:
    row = json.loads(next(f))
inputs = [Image.open(folder / p) for p in row["images"]]
steps, i = [], 0
while f"reasoning_thought_{i}" in row:
    image = row.get(f"reasoning_image_{i}")
    steps.append((row[f"reasoning_thought_{i}"], Image.open(folder / image) if image else None))
    i += 1
print(row["question"], row.get("options"), row["answer"])

Rows without images

text_only/ holds the rows of panoramic_completion whose images come from Matterport3D, and of exo_to_ego (Ego-Exo4D). 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.
Ego-Exo4D: obtain the dataset under its license at https://ego-exo4d-data.org/. Each image record names the take and the camera. The frame-selection code will be released with the training code.

Sources and licenses

Images keep the terms of the dataset they come from; check them before you use the data. Several sources allow non-commercial research use only. Licenses below are as stated by each source.

Task	Image sources	Terms
`bbox`	SA-1B; Visual-CoT (its images come from COCO, GQA, TextVQA, Visual Genome, Flickr30k, DocVQA, Open Images, CUB and others)	SA-1B Dataset Research License; Visual-CoT: CC BY-NC 4.0, images under their original terms
`crop`, `jigsaw_single_image`	SA-1B	SA-1B Dataset Research License
`jigsaw_multi_frame`	Hypersim, ScanNet, Open X-Embodiment robot datasets, Visual7W (COCO images) and other open images	Hypersim: CC BY-SA 3.0; ScanNet: ScanNet Terms of Use; Open X-Embodiment: CC BY 4.0; COCO images: Flickr terms
`object_segmentation`	LVIS v1 on COCO 2017 images	LVIS annotations: CC BY 4.0; COCO images: Flickr terms
`depth_map`, `marked_depth_map`	RefSpatial: Open Images, CA-1M and simulated scenes	RefSpatial: Apache-2.0; Open Images: CC BY 2.0; CA-1M: CC BY-NC-ND
`object_rotation`	MVImgNet	MVImgNet Terms of Use (non-commercial research)
`object_editing`	synthetic scenes in the style of CLEVR	CLEVR: CC BY 4.0
`birds_eye_view`, `birds_eye_view_rotation`	MindCube (its images come from ARKitScenes, WildRGB-D, DL3DV-10K and self-collected scenes)	MindCube: MIT; ARKitScenes: Apple license, non-commercial
`panoramic_completion`	Realsee3D, 360+x; Matterport3D (text only)	Realsee3D: CC BY 4.0; 360+x: CC BY-NC-SA 4.0; Matterport3D: Terms of Use
`object_viewpoint_render`, `object_viewpoint_render_oriented`, `object_orientation`	COCO 2017 images	Flickr terms (COCO terms of use)
`marked_position_view`	SAT (rendered rooms)	MIT
`exo_to_ego`	Ego-Exo4D (text only)	Ego-Exo4D license agreement

The reasoning text, the intermediate images and the release layout are ours. There is no blanket license over the images.