---
language:
- en
task_categories:
- visual-question-answering
tags:
- synthetic
- multi-image
- parquet
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---
# MultiImage Synthetic 440K

440,261 synthetic multi-image multiple-choice visual question-answering training examples, stored in 64 original Parquet shards with embedded JPEG image bytes (about 70.79 GB). These are the unchanged training Parquet shards from the scale_mix_v2 training mixture.

## Contents and loading

```python
from datasets import load_dataset
ds = load_dataset("yilingwang/MultiImage-Synthetic-440K", split="train", streaming=True)
example = next(iter(ds))
# example["images"] is a list of {"bytes": JPEG bytes, "path": synthetic image ID}.
```

Columns: `id`, `source_id`, `split`, `task_type`, `ability_id`, `sample_id`, `option_kind`, `split_groups`, `provenance`, `prompt`, `answer`, `images`, `image_sha256`.

`answer` is a JSON string containing choices and the correct option. `prompt` includes the question and image placeholders. `images` contains embedded JPEG bytes and synthetic image IDs; `image_sha256` contains their checksums. Original provenance strings are retained; internal source locations are historical references and are not needed to load the embedded images.

## Source and metadata audit

The source/export code identifies these examples as generated by the srcgen image generation, image editing, rendering, and synthetic video pipelines. No TikTok user-data ingestion path was found in this source chain. All 440,261 training records and 32 validation-stub records were checked: IDs and provenance match srcgen, and no TikTok identifiers, user-account field keys, or credential patterns were detected in the non-image columns. This audit checks sources and metadata; it does not constitute visual inspection of every image.

Generation and export code: https://github.com/yilinwangll/multi-image/tree/refactor-m0/srcgen

## Task distribution

| Task | Training examples |
|---|---:|
| action_order | 44804 |
| camera_turn | 43668 |
| change_match | 20231 |
| change_track | 20126 |
| count_across | 35514 |
| functional_match | 40751 |
| object_count | 9771 |
| object_group | 8382 |
| patch_fill | 38892 |
| point_match | 41978 |
| same_object | 11198 |
| same_person | 5693 |
| semantic_match | 44688 |
| spot_difference | 34027 |
| surround_views | 40538 |

## Validation stub and integrity

`validation_stub.parquet` contains 32 rows copied from the training mixture to satisfy the original trainer's file requirements. It is **not an independent evaluation set** and is deliberately excluded from the default dataset splits.

`SHA256SUMS` lists the SHA256 digest of every published file except itself. `source_manifest.json` is the original source manifest; its `train.parquet` hash refers to the original single-file source, while this repository publishes its 64 training shards. `audit_summary.json` summarizes the source and metadata checks.

The repository does not include model weights, optimizer states, raw user account records, or service credentials. No new license grant is asserted by this card.