ScaDSAI/tedrasim
收藏Hugging Face2026-03-19 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/ScaDSAI/tedrasim
下载链接
链接失效反馈官方服务：
资源简介：
---
license: mit
task_categories:
- image-to-text
- image-text-to-text
language:
- en
tags:
- synthetic
- multimodal
- scene-graph
- spatial-reasoning
- json
- computer-vision
- 3d
pretty_name: TEDRASIM Dataset
size_categories:
- 10K<n<100K
configs:
- config_name: default
drop_labels: true
---
# TEDRASIM Dataset
## Dataset Summary
This dataset is a multimodal training corpus for fine-tuning vision-language models to generate structured JSON scene-graph descriptions from rendered images.
This dataset contains:
- a synthetic dataset of 10,000 scenes, each rendered from 2 views (20,000 images total),
- a small real-world dataset of 120 images.
Each example consists of:
- one image showing a object assembly,
- a multi-turn chat-style prompt structure,
- a target JSON string describing the scene in a canonical structured format.
It is intended for research and development on structured visual reasoning, spatial reasoning, scene understanding, and image-to-JSON generation.
---
## Data Description
The dataset contains images of a very specific class of toy-like 3D objects.
These objects are:
- composed of geometric primitives such as cubes, sphere, cones,...
- arranged in simple spatial configurations
- rendered from multiple viewpoints
---
## Task Definition
The model is expected to generate a JSON scene graph describing the object by relative spatial relationships between a finite set of known fixed primitives
Relationships are defined locally between touching primitives, for example:
- "the blue cube is behind the green cone"
- "the red cube is left of the orange cylinder"
These relations are encoded explicitly in the JSON structure.
---
## Example Target Representation
A simplified example of a scene description:
```json
{
"primitive_counts": {
"red_cube": 2,
"yellow_sphere": 1
},
"primitives": [
{
"id": "P1",
"color": "red",
"shape": "cube",
"neighbors": {
"front": "P2",
"back": "Empty Space",
"left": "Empty Space",
"right": "Empty Space",
"up": "Empty Space",
"down": "Empty Space"
}
},
{
"id": "P2",
"color": "yellow",
"shape": "sphere",
"neighbors": {
"front": "Empty Space",
"back": "P1",
"left": "Empty Space",
"right": "Empty Space",
"up": "Empty Space",
"down": "Empty Space"
}
}
]
}
```
---
## Dataset Structure
### Data Instances
Each record in the JSONL files has the following structure:
```json
{
"image": "train/shard_0000/scene_000000_00.png",
"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "assistant", "content": "{\"primitive_counts\": ..., \"primitives\": ...}"}
],
"meta": {
"scene_id": "scene_000000",
"scene_hash": "...",
"split": "train",
"shard": "shard_0000",
"view_id": "view_00",
"num_primitives_in_scene": 4,
"min_primitives": 1,
"max_primitives": 6,
"seed": 42,
"attempt_index": 1,
"accepted_index": 0
}
}
```
### Data Fields
- image: relative path to the rendered image
- messages: chat-style training structure
- system: task instruction
- user: input prompt
- assistant: target JSON
- meta: auxiliary metadata for traceability
---
## Splits
The synthetic dataset is divided into:
- train.jsonl
- val.jsonl
- test.jsonl
The real dataset contains validation data only:
- val.jsonl
---
## Repository Layout
- synthetic/: synthetic dataset described here
- real/: real-world dataset component
- train.jsonl / val.jsonl / test.jsonl: split manifests
---
## License
MIT
提供机构：
ScaDSAI
