GLiNER UV Scripts

Zero-shot named-entity recognition over Hugging Face datasets using GLiNER. Pass a list of entity types at runtime — no fine-tuning required.

Script	What it does	Output
extract-entities.py	Extract entities from a text column with a custom set of types	New entities column (list of {start, end, text, label, score})

Quick start

Run on any HF dataset with a text column. No setup — uv resolves dependencies inline.

# Local CPU (small samples)
uv run extract-entities.py \
    librarian-bots/model_cards_with_metadata \
    yourname/model-cards-entities \
    --text-column card \
    --entity-types Person Organization Dataset Model Framework \
    --max-samples 100

On HF Jobs

# CPU job — fine for small/medium datasets, free or near-free
hf jobs uv run --flavor cpu-basic --secrets HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/gliner/raw/main/extract-entities.py \
    librarian-bots/model_cards_with_metadata \
    yourname/model-cards-entities \
    --text-column card \
    --entity-types Person Organization Dataset Model Framework \
    --max-samples 1000

# GPU job — worth it once you're processing >~1000 samples
hf jobs uv run --flavor t4-small --secrets HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/gliner/raw/main/extract-entities.py \
    librarian-bots/model_cards_with_metadata \
    yourname/model-cards-entities \
    --text-column card \
    --entity-types Person Organization Dataset Model Framework \
    --device cuda \
    --batch-size 32

Reading from local files or a mounted bucket

The input_dataset argument also accepts local file paths (parquet, jsonl, json, csv). Useful when the input is staged in a Storage Bucket — typical pattern for multi-stage pipelines where an upstream Job has prepared the data:

hf jobs uv run --flavor t4-small --secrets HF_TOKEN \
    -v hf://buckets/yourname/working-data:/input \
    https://huggingface.co/datasets/uv-scripts/gliner/raw/main/extract-entities.py \
    /input/data.parquet \
    yourname/output-entities \
    --text-column text --entity-types Person Organization Location \
    --device cuda --batch-size 32

Local paths are detected heuristically — anything starting with /, ./, ../, or ending in a known data extension is treated as a file path; otherwise the argument is interpreted as a HF dataset ID.

Recommended entity-type vocabularies

GLiNER is open-vocabulary, so any string works. Some starting points:

• General news/web text: Person Organization Location Date Event
• ML/AI text (e.g. model cards): Person Organization Dataset Model Framework Metric License
• Legal/policy: Person Organization Court Statute Date Jurisdiction
• Biomedical: Drug Disease Gene Protein Symptom

Quality drops on very abstract or polysemous types — start simple, iterate.

Models

Default: urchade/gliner_multi-v2.1 (multilingual, ~600 MB). Override with --gliner-model.

Other useful checkpoints:

• urchade/gliner_small-v2.1 — English, faster
• urchade/gliner_large-v2.1 — English, larger / higher quality
• knowledgator/gliner-multitask-large-v0.5 — multitask (NER + classification + relation)

See the Knowledgator org and urchade's models for the full set.

Pairing with Label Studio

Output of this script is a Hugging Face dataset of texts + extracted entities. To put those entities in front of human reviewers, see the bootstrap-labels skill (or the workflow it documents): pull this dataset's predictions into a Label Studio project for review, then export a corrected dataset back to the Hub.

Caveats

• GLiNER predictions are bootstrap labels — useful as a starting point, not as ground truth. Plan a review pass before downstream training.
• Texts longer than --max-text-chars (default 8000) are truncated. Long-form documents may need chunking + reassembly.
• Entity types are case-sensitive labels in output. Pass them as you want them to appear.