Shape of Thought (SoT-26K)

A large-scale dataset for progressive 3D object assembly with interleaved visual Chain-of-Thought traces. Accepted to ICML 2026.

---

Dataset Description

SoT-26K is a multimodal dataset containing 25,929 3D object assembly examples across 24 categories. Each example pairs a natural-language construction prompt with a step-by-step Shape-of-Thought reasoning trace, rendered intermediate assembly images, and a final completed object image.

The dataset is designed for interleaved vision-language training and evaluation: models can learn not only what final 3D object to produce, but also how to construct it through visually grounded intermediate states.

---

Dataset Structure

Each example in SoT-26K consists of:

• Prompt: a textual description of the target 3D object.
• Shape-of-Thought reasoning trace: a sequence of textual assembly steps with image placeholders such as [reasoning_image_1].
• Reasoning images: one or more rendered intermediate assembly states.
• Final assembly: a final textual completion marker with [final_image].
• Final image: the rendered completed object.

The main columns are:

• sample_id: source object identifier.
• category: object category/config name.
• prompt: natural-language construction prompt.
• shape_of_thought_reasoning_trace: interleaved step text with image placeholders.
• reasoning_image_1 ... reasoning_image_38: nullable rendered intermediate images.
• final_assembly: final assembly text with image placeholder.
• final_image: rendered final object image.

All image columns are stored as Hugging Face Image features, so the Dataset Viewer can render them directly.

---

Usage

To prepare interleaved text-image traces for training, replace each placeholder in shape_of_thought_reasoning_trace with the corresponding image column:

• [reasoning_image_1] maps to reasoning_image_1
• [reasoning_image_2] maps to reasoning_image_2
• [final_image] maps to final_image

For process-supervision training, each STEP n: segment can be treated as one progressive assembly step.

from datasets import load_dataset

dataset = load_dataset("yuhuo03/SoT-26K", "Chair", split="train")

example = dataset[0]
print(example["prompt"])
example["reasoning_image_1"]
example["final_image"]

---

Statistics

Property	Value
3D object samples	25,929
Object categories	24
Textual step annotations	162,003
Final assembly images	25,929
Splits	train, validation, test
Format	Hugging Face parquet with Image columns
License	CC-BY-4.0

Group	Category	Examples
Furniture & Infrastructure	Table	8,227
	Chair	6,323
	StorageFurniture	2,269
	Door	230
	Bed	192
Lifestyle & Accessories	Lamp	2,212
	Vase	1,086
	Hat	231
	Bag	126
	Scissors	68
Electronics & Office	Display	930
	Clock	554
	Laptop	433
	Earphone	229
	Keyboard	156
Kitchenware & Appliances	Faucet	648
	TrashCan	321
	Refrigerator	187
	Microwave	183
	Dishwasher	181
Dining & Tableware	Bottle	436
	Knife	327
	Mug	192
	Bowl	188
Total	24 categories	25,929

Statistics are detailed in the SoT paper.

---

Framework

---

Citation

If you use SoT-26K, please cite:

@misc{huo2026shapethoughtprogressiveobject,
      title={Shape of Thought: Progressive Object Assembly via Visual Chain-of-Thought}, 
      author={Yu Huo and Siyu Zhang and Kun Zeng and Haoyue Liu and Owen Lee and Junlin Chen and Yuquan Lu and Yifu Guo and Yaodong Liang and Xiaoying Tang},
      year={2026},
      eprint={2601.21081},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.21081}, 
}