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:

{
  "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:

{
  "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