eozlu's picture
Upload dataset
d873dfe verified
metadata
language:
  - en
license: apache-2.0
task_categories:
  - image-text-to-text
  - visual-question-answering
tags:
  - tangram
  - visual-reasoning
  - step-by-step
  - puzzle
  - colored
configs:
  - config_name: default
    data_files:
      - split: train_colored
        path: data/train_colored-*
      - split: test_colored
        path: data/test_colored-*
dataset_info:
  features:
    - name: id
      dtype: string
    - name: question_interleave
      list:
        - name: content
          dtype: string
        - name: index
          dtype: int64
        - name: type
          dtype: string
    - name: question_images
      list: image
    - name: solution_interleave
      list:
        - name: content
          dtype: string
        - name: index
          dtype: int64
        - name: type
          dtype: string
    - name: solution_images
      list: image
    - name: answer
      dtype: string
    - name: options
      dtype: 'null'
    - name: knowledge
      dtype: string
    - name: subknowledge
      dtype: string
  splits:
    - name: train_colored
      num_bytes: 56050665
      num_examples: 912
    - name: test_colored
      num_bytes: 6108402
      num_examples: 101
  download_size: 50824505
  dataset_size: 62159067

TangramData-Instruct-Colored

Step-by-step tangram assembly dataset with colored pieces and unique per-puzzle reference images.

Key Features

  • Each question includes a unique _square.png reference image showing the 7 tangram pieces arranged as a square with puzzle-specific coloring
  • Each step annotation includes remaining pieces list
  • Final step includes <answer> tags

Dataset Statistics

  • Train Samples: 912 (90%)
  • Test Samples: 101 (10%)
  • Total Samples: 1013
  • Question Images: 2 per sample (unique colored reference + target silhouette)
  • Solution Images: 7 per sample (assembly steps with colored pieces)

Format

This dataset follows the MathCanvas-Instruct interleaved format for compatibility with BAGEL-Canvas training pipelines.