empathyai/gliner_large-v2.5-books

Books Named Entity Recognition Model

This model specialises in recognising book titles and author names in short, user‑typed queries. It achieves > 92 % F1 on a held‑out evaluation set.

Check out this model in action in this experience!

---

1 Provenance

This model's provenance originates from fine-tuning gliner-community/gliner_large‑v2.5 using synthetic query data derived from Project Gutenberg's public catalogue.

---

2  Use‑case

This model is a drop‑in replacement for generic GLiNER when your text stream revolves around bibliographic requests such as:

“Looking for Dune from Frank Herbert.”
“Any recommendations by Mary Shelley?”

Typical applications:

• Query understanding in library / e‑book search engines
• Post‑processing LLM output to structure reading lists
• Digital humanities pipelines that need lightweight title/author extraction

Not suitable for: recognising publishers, ISBNs or long BIB‑style references (only short queries were used for training).

---

3  Performance

Metric	Overall	title	author
Precision	0.9999	0.9999	0.9999
Recall	0.8583	0.7661	0.9287
F1‑score	0.9237	0.8675	0.9630
Support	69 880	30 290	39 590

Evaluation dataset: 43 493 queries (English‑only) held out from the training corpus. Prediction threshold = 0.2.

---

4  Quick start

from gliner import GLiNER

model = GLiNER.from_pretrained("empathyai/gliner_large-v2.5-books")
text = "Looking for The Man in the High Castle by Philip K. Dick."

entities = model.predict_entities(text, ["title", "author"], threshold=0.2)
print(entities)
# [{'text': 'The Man in the High Castle', 'label': 'title', 'score': 0.99},
#  {'text': 'Philip K. Dick', 'label': 'author', 'score': 0.99}]

---

5  Training details

• Base model: gliner_large-v2.5 (≈ 459 M parameters, Creative Commons Zero v1.0 Universal)
• Dataset: empathyai/books-ner-dataset — 435 k synthetic English queries (titles + authors only)
• Splits: 391 432 train / 43 493 eval (duplicates removed)
• Script highlights
  • Learning rate 5 × 10⁻⁶, linear schedule, warm‑up 10 %
  • Batch 32, gradient accumulation 2, focal loss α 0.75 / γ 2
  • 1 epoch
  • Gradient checkpointing + BF16 for memory efficiency
  • Trained on a single L40S; total wall time ≈ 40 min

---

6  Limitations & bias

• The vocabulary of titles/authors comes from Project Gutenberg (public‑domain heavy; modern best‑sellers may be unseen).
• Only short, informal English queries were simulated. Long paragraphs or non‑English text may degrade accuracy.
• Does not tag publishers, dates, ISBNs, or other bibliographic fields.

---

7  Acknowledgements

Thanks to the GLiNER authors and maintainers; HuggingFace for hosting; Project Gutenberg volunteers for the free metadata.