storytracer/bhl_rights_json

Biodiversity Heritage Library Rights Metadata (JSON)

This dataset contains all unique combinations of the four rights-related metadata fields from the item.txt.gz file of the Biodiversity Heritage Library (BHL) open dataset on AWS Open Data. It is designed to be classified using the UV Dataset Classification Script to determine the public domain status of each item in the BHL collection.

Background

The BHL is a consortium of natural history libraries providing open access to biodiversity literature spanning from the 15th century to the present. The bulk of the collection is in the public domain, but it also contains in-copyright materials digitized with permission from rights holders.

The BHL Item table has four rights-related columns, but no single column reliably determines whether an item is in the public domain. This dataset combines all four fields into a single JSON string so they can be evaluated together by a language model classifier.

Fields

Field	Type	Description
CopyrightStatus	nvarchar(500)	Free-text copyright status from the BHL Item table.
RightsStatement	nvarchar(500)	Free-text rights statement from the BHL Item table.
LicenseType	nvarchar(200)	License URL or description. Mostly Creative Commons license URLs when present.
RightsHolder	nvarchar(255)	Name of the institution or person holding the rights.
rights_json	VARCHAR	JSON object combining the four fields above, with empty values set to null. This is the column to classify.
Count	INTEGER	Number of BHL items with this exact combination of rights fields.

Extraction Query

COPY (
  SELECT
    CopyrightStatus,
    RightsStatement,
    LicenseType,
    RightsHolder,
    CAST(to_json(
      {
        CopyrightStatus: NULLIF(CopyrightStatus, ''),
        RightsStatement: NULLIF(RightsStatement, ''),
        LicenseType: NULLIF(LicenseType, ''),
        RightsHolder: NULLIF(RightsHolder, '')
      }
    ) AS VARCHAR) AS rights_json,
    COUNT(*) AS Count
  FROM read_csv(
    'https://bhl-open-data.s3.amazonaws.com/data/item.txt.gz',
    delim='\t', header=true, all_varchar=true,
    nullstr='\N', quote='', escape=''
  )
  GROUP BY CopyrightStatus, RightsStatement, LicenseType, RightsHolder
  ORDER BY Count DESC
) TO 'bhl_rights_json.parquet' (FORMAT PARQUET);

Note: the CAST(... AS VARCHAR) ensures the rights_json column is stored as a plain UTF-8 string in Parquet, rather than a JSON/struct type that downstream tools may not parse correctly.

Classification

Task

The classification is a binary determination of public domain status:

• public_domain: Any positive public domain signal in any JSON field.
• not_public_domain: No positive public domain signal found in any JSON field.

Signal detection

All four fields are treated as equal sources of public domain signals. A single positive signal in any field is sufficient and definitive. The label descriptions provide illustrative examples of positive signals rather than an exhaustive list, to account for the wide variety of ways public domain status is expressed across contributing institutions:

• Phrases (in any field): "public domain", "not in copyright", "NIC" (Not In Copyright), "no known copyright restrictions", "no longer under copyright", "copyright expired", "out of copyright", "unaware of any copyright restrictions"
• License identifiers (in any field): Creative Commons Public Domain Mark, CC0
• Institutional assertions: Statements where an institution "believes", "considers", or "determines" that a work is not in copyright, treated as positive signals based on institutional due diligence

Creative Commons licenses other than Public Domain Mark and CC0 (such as CC-BY, CC-BY-SA, CC-BY-NC, CC-BY-NC-SA) are explicitly not public domain signals.

BHL Collection Policy

The BHL's Copyright and Reuse policy contains an explicit statement governing how public domain status interacts with other metadata:

As the Memorandum of Understanding signed by the Institutions composing BHL states, "All information currently in the public domain remains in the public domain. Neither BHL nor the data providers will seek to assert any IPR over public domain materials."

This means a public domain signal in any field cannot be overridden by the assertion of a license or other rights claim in another field. This principle is embedded in the label descriptions provided to the classifier.

Command

hf jobs uv run \
  --flavor l4x1 \
  --secrets HF_TOKEN \
  --timeout 1h \
  --with "vllm<0.12.0" \
  --with "flashinfer-python==0.5.2" \
  --with "flashinfer-cubin==0.5.2" \
  https://huggingface.co/datasets/uv-scripts/classification/raw/main/classify-dataset.py \
  --input-dataset "storytracer/bhl_rights_json" \
  --column "rights_json" \
  --model "Qwen/Qwen3-4B" \
  --labels "public_domain,not_public_domain" \
  --label-descriptions "public_domain:Any positive public domain signal in ANY JSON field is sufficient and definitive. The following are examples of positive signals and not an exhaustive list. They illustrate the variety of ways public domain status can be expressed. Any phrase that conveys the same meaning counts even if the exact wording differs or contains additional words. Example phrases: public domain, not in copyright, NIC (Not In Copyright), no known copyright restrictions, no longer under copyright, copyright expired, out of copyright, unaware of any copyright restrictions. Any mention of Creative Commons Public Domain Mark or CC0 in any field is also a positive public domain signal. Institutional assertions count as positive signals. If an institution believes or considers or determines that a work is not in copyright or is in the public domain that IS a positive signal based on due diligence. This data is from the Biodiversity Heritage Library (BHL) a consortium of natural history libraries. As the Memorandum of Understanding signed by the institutions composing BHL states all information currently in the public domain remains in the public domain and neither BHL nor the data providers will seek to assert any secondary intellectual property rights over public domain materials. Therefore a single public domain signal in any field asserts public domain status and cannot be overridden by any other field;not_public_domain:No positive public domain signal found in any JSON field. Note that Creative Commons licenses such as CC-BY or CC-BY-SA or CC-BY-NC or CC-BY-NC-SA are NOT public domain signals. Only the Creative Commons Public Domain Mark and CC0 are public domain signals" \
  --output-dataset "storytracer/bhl_rights_json_classified" \
  --enable-reasoning \
  --max-tokens 500

The vllm<0.12.0 constraint is required for compatibility with the UV classification script. The flashinfer-python==0.5.2 and flashinfer-cubin==0.5.2 pins resolve a version mismatch in the HF Jobs base image.

Results

The classified output is saved to storytracer/bhl_rights_json_classified.

Classification	Unique Combos	Items	Share
public_domain	260	203,297	63.5%
not_public_domain	1,466	117,090	36.5%

The output dataset includes a reasoning column documenting the model's step-by-step reasoning for each classification decision.

Related Resources

• BHL Open Data on AWS — source S3 bucket with scan images, OCR, and metadata
• BHL Open Data GitHub — documentation for the open data structure
• BHL Copyright and Reuse — BHL's copyright policy and MOU statement
• BHL Collection Development Policy — collection scope and copyright handling
• UV Dataset Classification Script — the classification tool