---
license: mit
pretty_name: Data Snapshot Metadata Gold
size_categories:
- n<1K
tags:
- metadata
- document-ai
- structured-data
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - prwp/figure/*.json
    - prwp/table/*.json
    - refugee/figure/*.json
    - refugee/table/*.json
    - unhcr/figure/*.json
    - unhcr/table/*.json
---

# Data Snapshot Metadata Gold

This dataset contains human-reviewed metadata for 102 data snapshots: figures
and tables extracted from institutional documents. Each JSON file describes one
snapshot image in the
[`ai4data/data-snapshot`](https://huggingface.co/datasets/ai4data/data-snapshot)
dataset.

This repository contains snapshot-level metadata only. It does not contain the
snapshot images or the separate source-document metadata published in
[`ai4data/data-snapshot`](https://huggingface.co/datasets/ai4data/data-snapshot).

## Dataset structure

All records are currently assigned to the `train` split.

| Source collection | Figures | Tables | Total |
| --- | ---: | ---: | ---: |
| UNHCR | 17 | 17 | 34 |
| PRWP | 17 | 17 | 34 |
| Refugee | 17 | 17 | 34 |
| **Total** | **51** | **51** | **102** |

The repository is organized by source collection and snapshot type:

```text
data_snapshot_metadata_gold/
├── prwp/{figure,table}/*.json
├── refugee/{figure,table}/*.json
├── unhcr/{figure,table}/*.json
├── data_snapshot_metadata_schema_v1.4.schema.json
└── README.md
```

Metadata filenames follow this convention:

```text
{source_document}_{figure|table}_{index}.json
```

The corresponding snapshot has the same basename and a `.png` extension in
[`ai4data/data-snapshot`](https://huggingface.co/datasets/ai4data/data-snapshot). Source-document names may contain underscores, so use
the full basename for matching rather than splitting on every underscore.

## Loading the dataset

```python
from datasets import load_dataset

dataset = load_dataset(
    "ai4data/data_snapshot_metadata_gold",
    split="train",
)
```

## Metadata schema

The records conform to **Data Snapshot Metadata Schema v1.4**. The included
[`data_snapshot_metadata_schema_v1.4.schema.json`](data_snapshot_metadata_schema_v1.4.schema.json)
is a frozen, generated JSON Schema for this dataset release.

The canonical schema is implemented in Pydantic. This release is pinned to
commit
[`2b317a4d95b1c929b32c790b678646fdfca16eef`](https://github.com/ajdajd/data-snapshot/tree/2b317a4d95b1c929b32c790b678646fdfca16eef/src/data_snapshot/metadata_schema).

The JSON Schema uses JSON Schema Draft 2020-12. The complete field definitions,
constraints, descriptions, and examples are provided in the schema file.

Schema v1.4 is fixed for this gold release. Subsequent schema changes will use
a new version, beginning with v1.5, and will require a separate metadata review
and dataset update.

## Dataset creation

The metadata was initially produced through model-assisted structured
extraction from data snapshot images. The generated records were subsequently
reviewed and corrected by a human annotator against the images. Source-document
metadata was not supplied to the extraction model.

"Gold" means human-reviewed. It does not imply that every field is objectively
unambiguous or independently verified against the complete source document.

All 102 records were validated against the canonical Schema v1.4 Pydantic
model before publication.

## Intended uses

The dataset supports work on:

- structured metadata extraction from figures and tables;
- evaluation of schema-constrained extraction systems;
- metadata quality review and error analysis; and
- research on describing data-bearing content in institutional documents.

## Limitations

- The dataset contains 102 selected snapshots and is not representative of all
  institutional documents or visualization types.
- Metadata is based on evidence visible in each snapshot image. Relevant
  context elsewhere in the source document may not be represented.
- Some fields require interpretation and may admit more than one reasonable
  annotation.
- The current release contains only a `train` split. Train/test assignments
  will be defined in a later dataset release.

## License

The metadata in this repository is released under the MIT License. Snapshot
images are not included here and remain subject to the licensing and
attribution terms of their source dataset and original publishers.

## Paper and citation

[TBD]