davanstrien/eval-mentions-bootstrap-v2

Bootstrap NER dataset produced by urchade/gliner_multi-v2.1 over /input/cleaned-cards-quality.parquet.

Generated using uv-scripts/gliner/extract-entities.py.

Provenance

Source dataset	/input/cleaned-cards-quality.parquet (split train)
Text column	card
Bootstrap model	urchade/gliner_multi-v2.1
Entity types	benchmark name, evaluation metric
Confidence threshold	0.6
Samples processed	5000
Total entities extracted	4492
Inference device	cuda
Wall clock	553.9s (9.03 samples/s)

Schema

Original /input/cleaned-cards-quality.parquet columns plus an entities column:

entities: list of {
    "start": int,    # character offset, inclusive
    "end": int,      # character offset, exclusive
    "text": str,     # the matched span
    "label": str,    # one of ['benchmark name', 'evaluation metric']
    "score": float,  # GLiNER confidence in [0, 1]
}

Caveats

• These are bootstrap labels, not human-reviewed. Treat low-confidence (< 0.7) entities as candidates for review.
• GLiNER is zero-shot: changing --entity-types changes what it extracts, but quality varies by entity type.
• Long texts were truncated at 8000 characters before inference.