Datasets:

cometadata
/

crossref-datacite-citations-datacite-enrichment-format

---
license: cc0-1.0
pretty_name: Crossref citations of DataCite DOIs in DataCite enrichment format
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.jsonl.gz
tags:
  - datacite
  - crossref
  - citations
  - metadata
  - related-identifiers
size_categories:
  - 1M<n<10M
---
# Crossref citations of DataCite DOIs in DataCite enrichment format

This dataset holds citation insertions derived from the `all` configuration of
[cometadata/crossref-datacite-citations](https://huggingface.co/datasets/cometadata/crossref-datacite-citations).
Every JSONL record follows the [COMET enrichment input schema](https://github.com/cometadata/comet-enrich/blob/abd6a9acd3fc70d25f703ea5be3c0fc62e0b6a61/configs/enrichment_input_schema.json),
with an `enrichedValue` following [DataCite RelatedIdentifier 4.7](https://datacite-metadata-schema.readthedocs.io/en/4.7/properties/relatedidentifier/).

Each record targets a cited DataCite DOI in `doi`. It inserts an `IsCitedBy` relation to the DOI of the Crossref work that cites it.

`sourceId` is the project DOI `10.82461/xg8p-mk16` in every record. Earlier revisions of this dataset had an empty `sourceId`. The content key excludes `sourceId`, so each `contentKey` is the same as in those revisions.

## Contents

The dataset contains **7,339,071 citations of 1,705,826 DataCite DOIs**, stored in 30 gzip-compressed JSONL shards.

| Measure | Count |
| --- | ---: |
| Source rows and unique cited DOIs | 1,705,826 |
| Cited DOIs with output records | 1,705,826 |
| Source citing-work entries | 7,339,071 |
| Source reference matches | 7,385,186 |
| Output records | 7,339,071 |
| Records whose citing work's strongest provenance is `publisher` | 1,192,204 |
| Records whose citing work's strongest provenance is `crossref` | 77,601 |
| Records whose citing work's strongest provenance is `mined` | 6,069,266 |
| Records whose cited DOI is an arXiv DOI | 5,999,691 |

Each source row lists every citing work once, so each pair of cited DOI and citing DOI yields one record. A citing work can match the cited DOI in more than one of its references. That is why source reference matches exceed records.

The source gives each citing-work entry the strongest provenance among its matches, ranked `publisher`, then `crossref`, then `mined`. The provenance rows above count that value for information. The records do not carry it.

## Mapping

- Each record uses `action: "insert"` and `field: "relatedIdentifiers"`. `originalValue` is absent.
- The source `doi` maps to the record's `doi`. It is the cited DataCite DOI.
- Each `cited_by[].doi` maps to `enrichedValue.relatedIdentifier`. The value is a bare DOI name, as in the DataCite 4.7 relatedIdentifier example, not a `https://doi.org/` URL.
- `relatedIdentifierType` is always `DOI`. `relationType` is always `IsCitedBy`.
- `resourceTypeGeneral` is omitted because the source does not record the citing work's type.
- The records do not carry the source match details (`match_type`, `raw_match`, and `reference`) or the citing work's `provenance`.
- Both DOIs must pass the comet-enrich `is_valid_doi_name` check and be lowercase. A violation stops the conversion instead of dropping the record. The pinned source has none.
- A repeated pair of cited DOI and citing DOI stops the conversion. The pinned source has none.

DOIs are copied as the source supplies them, already lowercase. Four cited DOIs contain non-ASCII characters, for example `10.13125/américacrítica/3012`. The JSONL files store them as UTF-8, not as `\u` escapes.

## Example

```json
{
  "doi": "10.48550/arxiv.1412.6980",
  "action": "insert",
  "field": "relatedIdentifiers",
  "enrichedValue": {
    "relatedIdentifier": "10.1001/jamadermatol.2022.5415",
    "relatedIdentifierType": "DOI",
    "relationType": "IsCitedBy"
  },
  "sourceId": "10.82461/xg8p-mk16",
  "contentKey": "63f52c05c46e67981b4f1f5157606c91"
}
```

## Content keys

The method name is `crossref-datacite-citations`. Each `contentKey` is the 32-character lowercase hexadecimal XXH3-128 hash of the RFC 8785 JCS encoding of:

```text
["crossref-datacite-citations", doi, "relatedIdentifiers", "insert", enrichedValue]
```

This follows the current [COMET content-key implementation](https://github.com/cometadata/comet-enrich/blob/abd6a9acd3fc70d25f703ea5be3c0fc62e0b6a61/crates/core/src/content_key.rs).
`sourceId` is not part of the key. The method name, the source ID, and all input and output file checksums are recorded in `manifest.json`.

## Use

Download the files under `data/` for direct enrichment ingestion. Each decompressed line is one schema-valid insertion record. The records do not carry provenance. Readers who want only citations whose DOI the publisher or Crossref asserted can use the `asserted` configuration of the source dataset.

These are proposed insertions. This dataset does not apply them to DataCite records. It does not compare them with the relatedIdentifiers already deposited in DataCite, so some records may duplicate an existing relation.

The `default` configuration has one `train` split. Every record has all three `enrichedValue` properties, so loading through Arrow adds no null properties.

## Reproduce and validate

Requires Python 3.11 or newer, the packages in `requirements.txt`, and an installed `hf` CLI.

```bash
python -m pip install -r requirements.txt
hf download cometadata/crossref-datacite-citations \
  --type dataset \
  --revision 65642a869543640f2965bcf9a7652f49fe88ef93 \
  --include 'data/all/*.parquet' \
  --include 'metadata/pipeline-manifest.json' \
  --local-dir source
python convert.py --source-dir source --output-dir rebuilt --source-id 10.82461/xg8p-mk16
cp provenance/enrichment_input_schema.json rebuilt/provenance/
cp provenance/independent_projection.sql rebuilt/provenance/
python validate.py rebuilt
python verify_source.py rebuilt --source-dir source/data/all
```

`convert.py` checks the SHA-256 of the pinned pipeline manifest, then the size and SHA-256 of each `all` parquet file that the manifest lists, before it reads any rows. It validates every emitted record against the pinned enrichment schema and a related-identifier subset schema. It also checks its source totals against the counts in the pipeline manifest. `validate.py` reads all output shards and checks their SHA-256 checksums, both schemas, and both DOIs. It independently recomputes content keys and checks global key uniqueness and the manifest totals. `verify_source.py` compares every output pair of cited DOI and citing DOI with a separate SQL projection of the source, in both directions, and checks that the counts match. Results are in `validation_report.json`. Two conversions of the pinned source produced byte-identical data files.

The source revision is `65642a869543640f2965bcf9a7652f49fe88ef93`, built by extractor version `2.2.0` from the Crossref snapshot and DataCite release `2026-07`. The COMET schema revision is `abd6a9acd3fc70d25f703ea5be3c0fc62e0b6a61`. `provenance/` holds the schema snapshot, the related-identifier subset schema, the source pipeline manifest, the independent SQL projection, and the first four records.

## Limitations

Citations reflect one Crossref snapshot and one DataCite release and keep the source's matching errors. The source notes that `mined` matches come from noisy reference text and may be lower quality. In 6,069,266 records, the citing work's strongest provenance is `mined`. In 5,999,691 records (81.75%), the cited DOI is an arXiv DOI with the prefix `10.48550/`. This conversion does not check the relatedIdentifiers already deposited in DataCite, and it does not record the citing work's resource type. The source is released under CC0 1.0, and so is this conversion.

	---
	license: cc0-1.0
	pretty_name: Crossref citations of DataCite DOIs in DataCite enrichment format
	configs:
	- config_name: default
	data_files:
	- split: train
	path: data/*.jsonl.gz
	tags:
	- datacite
	- crossref
	- citations
	- metadata
	- related-identifiers
	size_categories:
	- 1M<n<10M
	---

	# Crossref citations of DataCite DOIs in DataCite enrichment format

	This dataset holds citation insertions derived from the `all` configuration of
	[cometadata/crossref-datacite-citations](https://huggingface.co/datasets/cometadata/crossref-datacite-citations).
	Every JSONL record follows the [COMET enrichment input schema](https://github.com/cometadata/comet-enrich/blob/abd6a9acd3fc70d25f703ea5be3c0fc62e0b6a61/configs/enrichment_input_schema.json),
	with an `enrichedValue` following [DataCite RelatedIdentifier 4.7](https://datacite-metadata-schema.readthedocs.io/en/4.7/properties/relatedidentifier/).

	Each record targets a cited DataCite DOI in `doi`. It inserts an `IsCitedBy` relation to the DOI of the Crossref work that cites it.

	`sourceId` is the project DOI `10.82461/xg8p-mk16` in every record. Earlier revisions of this dataset had an empty `sourceId`. The content key excludes `sourceId`, so each `contentKey` is the same as in those revisions.

	## Contents

	The dataset contains 7,339,071 citations of 1,705,826 DataCite DOIs, stored in 30 gzip-compressed JSONL shards.

	\| Measure \| Count \|
	\| --- \| ---: \|
	\| Source rows and unique cited DOIs \| 1,705,826 \|
	\| Cited DOIs with output records \| 1,705,826 \|
	\| Source citing-work entries \| 7,339,071 \|
	\| Source reference matches \| 7,385,186 \|
	\| Output records \| 7,339,071 \|
	\| Records whose citing work's strongest provenance is `publisher` \| 1,192,204 \|
	\| Records whose citing work's strongest provenance is `crossref` \| 77,601 \|
	\| Records whose citing work's strongest provenance is `mined` \| 6,069,266 \|
	\| Records whose cited DOI is an arXiv DOI \| 5,999,691 \|

	Each source row lists every citing work once, so each pair of cited DOI and citing DOI yields one record. A citing work can match the cited DOI in more than one of its references. That is why source reference matches exceed records.

	The source gives each citing-work entry the strongest provenance among its matches, ranked `publisher`, then `crossref`, then `mined`. The provenance rows above count that value for information. The records do not carry it.

	## Mapping

	- Each record uses `action: "insert"` and `field: "relatedIdentifiers"`. `originalValue` is absent.
	- The source `doi` maps to the record's `doi`. It is the cited DataCite DOI.
	- Each `cited_by[].doi` maps to `enrichedValue.relatedIdentifier`. The value is a bare DOI name, as in the DataCite 4.7 relatedIdentifier example, not a `https://doi.org/` URL.
	- `relatedIdentifierType` is always `DOI`. `relationType` is always `IsCitedBy`.
	- `resourceTypeGeneral` is omitted because the source does not record the citing work's type.
	- The records do not carry the source match details (`match_type`, `raw_match`, and `reference`) or the citing work's `provenance`.
	- Both DOIs must pass the comet-enrich `is_valid_doi_name` check and be lowercase. A violation stops the conversion instead of dropping the record. The pinned source has none.
	- A repeated pair of cited DOI and citing DOI stops the conversion. The pinned source has none.

	DOIs are copied as the source supplies them, already lowercase. Four cited DOIs contain non-ASCII characters, for example `10.13125/américacrítica/3012`. The JSONL files store them as UTF-8, not as `\u` escapes.

	## Example

	```json
	{
	"doi": "10.48550/arxiv.1412.6980",
	"action": "insert",
	"field": "relatedIdentifiers",
	"enrichedValue": {
	"relatedIdentifier": "10.1001/jamadermatol.2022.5415",
	"relatedIdentifierType": "DOI",
	"relationType": "IsCitedBy"
	},
	"sourceId": "10.82461/xg8p-mk16",
	"contentKey": "63f52c05c46e67981b4f1f5157606c91"
	}
	```

	## Content keys

	The method name is `crossref-datacite-citations`. Each `contentKey` is the 32-character lowercase hexadecimal XXH3-128 hash of the RFC 8785 JCS encoding of:

	```text
	["crossref-datacite-citations", doi, "relatedIdentifiers", "insert", enrichedValue]
	```

	This follows the current [COMET content-key implementation](https://github.com/cometadata/comet-enrich/blob/abd6a9acd3fc70d25f703ea5be3c0fc62e0b6a61/crates/core/src/content_key.rs).
	`sourceId` is not part of the key. The method name, the source ID, and all input and output file checksums are recorded in `manifest.json`.

	## Use

	Download the files under `data/` for direct enrichment ingestion. Each decompressed line is one schema-valid insertion record. The records do not carry provenance. Readers who want only citations whose DOI the publisher or Crossref asserted can use the `asserted` configuration of the source dataset.

	These are proposed insertions. This dataset does not apply them to DataCite records. It does not compare them with the relatedIdentifiers already deposited in DataCite, so some records may duplicate an existing relation.

	The `default` configuration has one `train` split. Every record has all three `enrichedValue` properties, so loading through Arrow adds no null properties.

	## Reproduce and validate

	Requires Python 3.11 or newer, the packages in `requirements.txt`, and an installed `hf` CLI.

	```bash
	python -m pip install -r requirements.txt
	hf download cometadata/crossref-datacite-citations \
	--type dataset \
	--revision 65642a869543640f2965bcf9a7652f49fe88ef93 \
	--include 'data/all/*.parquet' \
	--include 'metadata/pipeline-manifest.json' \
	--local-dir source
	python convert.py --source-dir source --output-dir rebuilt --source-id 10.82461/xg8p-mk16
	cp provenance/enrichment_input_schema.json rebuilt/provenance/
	cp provenance/independent_projection.sql rebuilt/provenance/
	python validate.py rebuilt
	python verify_source.py rebuilt --source-dir source/data/all
	```

	`convert.py` checks the SHA-256 of the pinned pipeline manifest, then the size and SHA-256 of each `all` parquet file that the manifest lists, before it reads any rows. It validates every emitted record against the pinned enrichment schema and a related-identifier subset schema. It also checks its source totals against the counts in the pipeline manifest. `validate.py` reads all output shards and checks their SHA-256 checksums, both schemas, and both DOIs. It independently recomputes content keys and checks global key uniqueness and the manifest totals. `verify_source.py` compares every output pair of cited DOI and citing DOI with a separate SQL projection of the source, in both directions, and checks that the counts match. Results are in `validation_report.json`. Two conversions of the pinned source produced byte-identical data files.

	The source revision is `65642a869543640f2965bcf9a7652f49fe88ef93`, built by extractor version `2.2.0` from the Crossref snapshot and DataCite release `2026-07`. The COMET schema revision is `abd6a9acd3fc70d25f703ea5be3c0fc62e0b6a61`. `provenance/` holds the schema snapshot, the related-identifier subset schema, the source pipeline manifest, the independent SQL projection, and the first four records.

	## Limitations

	Citations reflect one Crossref snapshot and one DataCite release and keep the source's matching errors. The source notes that `mined` matches come from noisy reference text and may be lower quality. In 6,069,266 records, the citing work's strongest provenance is `mined`. In 5,999,691 records (81.75%), the cited DOI is an arXiv DOI with the prefix `10.48550/`. This conversion does not check the relatedIdentifiers already deposited in DataCite, and it does not record the citing work's resource type. The source is released under CC0 1.0, and so is this conversion.