README.md

---
configs:
- config_name: combined_eval5
  default: true
  data_files:
  - split: test
    path:
    - parquet/section1/test-*.parquet
    - parquet/section2/test-*.parquet
- config_name: section1_eval5
  data_files:
  - split: test
    path: parquet/section1/test-*.parquet
- config_name: section2_eval5
  data_files:
  - split: test
    path: parquet/section2/test-*.parquet
task_categories:
- visual-question-answering
tags:
- synthetic
- visual-question-answering
- evaluation
- parquet
---

# BlindLoop Evaluation

The frozen evaluation cohorts for the BlindLoop paper. Each eligible generated
task contributes exactly five deterministic, pixel-distinct image instances.
The five rows share a task's selected question/prompt family while varying the
rendered scene and gold answer as determined by the task's pixel oracle.

| Config | Tasks | Rows | Documented exclusions |
|---|---:|---:|---:|
| `section1_eval5` | 1,298 | 6,490 | 3 |
| `section2_eval5` | 874 | 4,370 | 1 |
| `combined_eval5` | 2,172 | 10,860 | 4 |

Each row exposes a native `image`, `question`, `answer`, `answer_options`, stable
`evaluation_id`, `instance_rank`, image hash, generator provenance, profile, and
verification metadata. Section 2 rows additionally expose their pixel-diversity
cell; those columns are empty for Section 1.

## Gold-label safety

The dataset includes `answer` for scoring. A model request must use only `image`,
`question`, and `answer_options`; do not serialize the full row into the prompt.
For stricter orchestration, use the gold-free and gold-only CSV ledgers under
`ledgers/`, whose source hashes are frozen in `manifests/source-sections.json`.

```python
from datasets import load_dataset

ds = load_dataset("taesiri/BlindLoop-Evaluation", "section1_eval5", split="test")
row = ds[0]
row["image"].show()
print(row["question"], row["answer_options"], row["answer"])
```