Datasets:

tasksource
/

tron-mcq

dataset_info:
  features:
    - name: images
      list:
        image:
          decode: false
    - name: image_group_id
      dtype: string
    - name: qa
      list:
        - name: inputs
          dtype: string
        - name: choices_list
          list: string
        - name: labels
          dtype: int64
        - name: metadata
          dtype: string
  splits:
    - name: train
      num_bytes: 5769627677
      num_examples: 94157
    - name: validation
      num_bytes: 4020063
      num_examples: 83
  download_size: 5773110820
  dataset_size: 5773647740
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
language:
  - en
license: other
source_datasets: []
task_categories:
  - multiple-choice
  - visual-question-answering
tags:
  - decision-model
  - procedural
license_name: tron-research-use
license_link: https://github.com/YangTianze009/TRON#license

tron-mcq

TRON MC keeps native images, options and labels, clarifying rounded measurements.

Source: https://github.com/YangTianze009/TRON (research-use terms). Some geometry figures display rounded measurements while their source gold retains additional precision. Replace the exact two-decimal output request with a closest-offered-value MC instruction. Original questions remain in metadata. This repairs the pinned 100,000-QA mirror without regenerating any images or changing supervision.

Original data: original distribution. Repackaged as parquet for tasksource by scripts/repackage_dataset/procedural/tron.py.

Conversion provenance: source revisions, hashes, configuration and exclusions.

Generation checkpoints record each chunk’s seeds and exclusions. Reproduce with the resumable scale script.