---
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
---

# PixelProof Evaluation

The frozen evaluation sets for the PixelProof paper. Each eligible
replay-verified world contributes exactly five deterministic, pixel-distinct
instances. The five instances of a world share its selected question text but show different rendered scenes; the answer can differ between instances, and the world's inverse program recovers it from the image alone.

| Config | Experiment | Worlds | Instances | Documented exclusions |
|---|---|---:|---:|---:|
| `section1_eval5` | Profile-steered generation | 1,298 | 6,490 | 3 |
| `section2_eval5` | Spatially-steered generation | 874 | 4,370 | 1 |
| `combined_eval5` | Both experiments | 2,172 | 10,860 | 4 |

The config names, the `section` column, and the `ledgers/section*` files are
stable identifiers: `section1` holds profile-steered worlds and `section2`
holds spatially-steered worlds. The exclusions are the worlds that passed the
200-scene replay (1,301 profile-steered and 875 spatially-steered) but
produced fewer than five pixel-distinct scenes; each one is documented in the
corresponding `ledgers/section*_exclusions.csv` file.

Each row exposes a native `image`, `question`, `answer`, `answer_options`, a
stable `evaluation_id`, `instance_rank`, an image hash, coding-agent provenance,
the discovery profile, and verification metadata. Spatially-steered rows also
record their spatial pattern (`pixel_support_scale`, `pixel_support_shape`,
and `pixel_diversity_cell`), one of nine patterns that combine a scale
(focal, regional, distributed) with a shape (compact, pathlike, multipart).
These columns are empty for profile-steered rows.

## Keeping answers out of prompts

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 answer-free and answer-only CSV ledgers under `ledgers/`, whose source hashes are frozen in
`manifests/source-sections.json`.

```python
from datasets import load_dataset

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

## Release integrity

This release preserves the questions, answers, images, and dataset
configurations used in the paper. Local filesystem paths in audit diagnostics
are redacted, and `CHECKSUMS.sha256` lists SHA-256 checksums for the files
distributed here.

## Project resources

- [PixelProof website](https://projectpixelproof.github.io/)
- [Source code](https://github.com/ProjectPixelProof/PixelProof)