yunfengwang/TVP-Training-Data

TVP Training Data — Thinking with Visual Primitives

Training data for the Thinking with Visual Primitives PyTorch implementation.

Overview

This dataset contains all training data for the multi-stage TVP pipeline:

Split	File	Samples	Description
Pretrain	pretrain/grounding.jsonl	146K	COCO-based grounding (label + bbox)
SFT	sft/grounding/sft_grounding.jsonl	30K	Grounding with structured thinking + negatives (15%)
SFT	sft/counting/counting_data.jsonl	8K	Counting with bbox grounding in CoT
SFT	sft/spatial/spatial_data.jsonl	3K	CLEVR-style spatial reasoning
SFT	sft/maze/maze_data.jsonl	5K	Procedural maze navigation (point primitives)
SFT	sft/path/path_data.jsonl	3K	Path tracing (point sequences)

Data Format

All files are JSONL. Coordinates are normalized integers in [0, 999].

Pretrain Grounding

{
  "image": "images/000000000009.jpg",
  "label": "person",
  "boxes": [[480, 201, 720, 850]],
  "points": [],
  "normalized": true
}

SFT Grounding (with structured thinking)

{
  "image": "images/000000000009.jpg",
  "question": "Locate the person in the image.",
  "thinking": "1. **Analyzing the request**\nThe user asks me to locate the person in this image.\n2. **Object grounding**\nI see a <|ref|>person<|/ref|><|box|>[[480,201,720,850]]<|/box|>.\n3. **Conclusion**\nThe person is located at the specified coordinates.",
  "answer": "The person is located at [[480,201,720,850]].",
  "boxes": [[480, 201, 720, 850]],
  "points": []
}

SFT Counting

{
  "image": "images/000000000025.jpg",
  "question": "How many people are in this image?",
  "thinking": "1. **Analyzing the request**\nThe user asks me to count the person in this image.\n2. **Object grounding**\nI see 2 instance(s) of <|ref|>person<|/ref|><|box|>[[338,121,630,923],[634,154,888,945]]<|/box|>.\n3. **Conclusion**\nThere are 2 person in this image.",
  "count": 2,
  "boxes": [[338, 121, 630, 923], [634, 154, 888, 945]]
}

Maze / Path (point primitives)

{
  "image": "images/maze_00001.png",
  "question": "Navigate from start to end in this maze.",
  "thinking": "... DFS exploration with <|point|>[[x,y]]<|/point|> waypoints ...",
  "answer": "...",
  "points": [[100, 200], [150, 250], [200, 300]]
}

Visual Primitives

# Bounding box
<|ref|>cat<|/ref|><|box|>[[x1,y1,x2,y2]]<|/box|>

# Multiple boxes
<|ref|>person<|/ref|><|box|>[[130,50,400,800],[500,60,750,790]]<|/box|>

# Point sequence
<|point|>[[100,200],[150,250],[200,300]]<|/point|>

Generation Scripts

The scripts/ folder contains all data generation code:

Script	Purpose
prepare_all_data.py	One-command pipeline (downloads COCO + generates all data)
generate_sft_grounding_data.py	Grounding with negatives + diverse prompt templates
generate_maze_data.py	Procedural maze generation with DFS solutions
generate_path_data.py	Path tracing data generation

Regenerate from scratch

# Full pipeline (downloads COCO 2017 val ~1GB)
python scripts/prepare_all_data.py \
    --output_dir data --coco_split val --coco_subset 5000

# Generate grounding with negatives
python scripts/generate_sft_grounding_data.py \
    --coco_jsonl data/pretrain/grounding.jsonl \
    --image_root data/coco/val \
    --output data/sft/grounding/sft_grounding.jsonl \
    --neg_ratio 0.15 --max_samples 30000

Source Images

The JSONL files reference COCO 2017 images. Download them separately:

• Train: COCO 2017 Train (18GB)
• Val: COCO 2017 Val (1GB)

For maze/spatial/path tasks, images are procedurally generated by the scripts.

Related

• GitHub Repository - Full training code and pipeline
• TVP-OPD-Qwen2VL-2B — Final distilled model
• TVP-SFTBox-Qwen2VL-2B — Box expert
• TVP-SFTPoint-Qwen2VL-2B — Point expert
• TVP-Pretrain-Qwen2VL-2B — Pretrained base

Citation

@software{wang2026tvp_pytorch,
  title={Thinking with Visual Primitives — PyTorch Implementation},
  author={Wang, Weishan},
  url={https://github.com/vra/Thinking-with-Visual-Primitives-pytorch},
  year={2026}
}

License

MIT