Datasets:

taesiri
/

BlindLoop-Generations

configs:
  - config_name: combined_all
    default: true
    data_files:
      - split: train
        path:
          - parquet/section1/train-*.parquet
          - parquet/section2/train-*.parquet
  - config_name: section1_all
    data_files:
      - split: train
        path: parquet/section1/train-*.parquet
  - config_name: section2_all
    data_files:
      - split: train
        path: parquet/section2/train-*.parquet
task_categories:
  - visual-question-answering
tags:
  - synthetic
  - visual-question-answering
  - parquet

BlindLoop Generations

A flat, general-purpose visual-question-answering dataset generated by the BlindLoop paper experiments. Each row is one concrete question instance with a native Hugging Face Image value, question, gold answer, answer choices, and fully filterable generation provenance.

Config	Rows	Tasks	Unique source images
`section1_all`	516,810	1,301	249,488
`section2_all`	335,271	875	166,364
`combined_all`	852,081	2,176	415,852

Main columns

image: embedded image bytes decoded by Hugging Face as an image.
question, answer, answer_options: the VQA instance and gold target.
profile, generator_model, generator_harness, reasoning_effort: generation treatment.
campaign_id, candidate_id, record_id, example_id, prompt_family: exact provenance.
margin, quarantined, mechanically_eligible, human_admission: verification and review state.
pixel_support_scale, pixel_support_shape, pixel_diversity_cell: Section 2 pixel-diversity descriptors; empty for Section 1.
image_sha256, sample_id: immutable media and row identifiers.

This archive includes all verified replay outputs, including explicitly flagged quarantined rows for auditability. Mechanical verification is not a substitute for human admission; human_admission preserves the source status.

from datasets import load_dataset

ds = load_dataset("taesiri/BlindLoop-Generations", "section1_all", split="train")
row = ds[0]
row["image"].show()
print(row["question"], row["answer"])