Pretrained Models

This page provides detailed information about the models used in GLiNKER entity linking pipelines. GLiNKER uses separate models for different pipeline layers: NER (L1), Linking/Disambiguation (L3), and Reranking (L4).

NER Models (L1)

NER models detect entity mentions in the input text. GLiNKER uses GLiNER bi-encoder models for zero-shot NER with support for arbitrary entity types. BiEncoder variants allow precomputing label embeddings once and reusing them across millions of documents, delivering significant speedups for large-scale linking.

Name▼	Parameters▼	Text Encoder▼	Label Encoder▼	Avg. CrossNER▼	Speed H100 (ex/s)▼	Speed Pre-computed (ex/s)▼
gliner-bi-large-v2.0	530M	ettin-encoder-400m	bge-base-en-v1.5	61.5%	2.68	3.6
gliner-bi-base-v2.0	194M	ettin-encoder-150m	bge-small-en-v1.5	60.3%	5.91	9.51
gliner-bi-small-v2.0	108M	ettin-encoder-68m	all-MiniLM-L12-v2	57.2%	7.99	15.22
gliner-bi-edge-v2.0	60M	ettin-encoder-32m	all-MiniLM-L6-v2	54%	13.64	24.62

Basic Usage

from gliner import GLiNER

model = GLiNER.from_pretrained("knowledgator/gliner-bi-base-v2.0")

text = """
Cristiano Ronaldo dos Santos Aveiro was born on 5 February 1985 in Portugal.
He plays for Al Nassr and has won five Ballon d'Or awards.
"""

labels = ["person", "award", "date", "teams", "location"]

entities = model.predict_entities(text, labels, threshold=0.3)

for entity in entities:
    print(entity["text"], "=>", entity["label"])

Expected Output

Cristiano Ronaldo dos Santos Aveiro => person
5 February 1985 => date
Portugal => location
Al Nassr => teams
Ballon d'Or => award

Pre-compute Labels Embeddings

labels = ["person", "award", "date", "teams", "location"]

entity_embeddings = model.encode_labels(labels, batch_size=8)

output = model.batch_predict_with_embeds([text], entity_embeddings, labels)
for entities in output:
  for entity in entities:
      print(entity["text"], "=>", entity["label"])

Expected Output

Encoding labels: 100%|██████████| 1/1 [00:00<00:00,  2.51it/s]
Cristiano Ronaldo dos Santos Aveiro => person
5 February 1985 => date
Portugal => location
Al Nassr => teams
Ballon d'Or => award

Linking Models (L3)

Linking models perform entity disambiguation by computing similarity between mention contexts and candidate entity descriptions. These models use a cross-encoder architecture based on DeBERTa.

Model▲	Base Encoder▲	Use Case▲
gliner-linker-base-v1.0	deberta-base	Balanced performance
gliner-linker-large-v1.0	deberta-large	Maximum accuracy

Usage in Pipeline

from glinker import ConfigBuilder, DAGExecutor

builder = ConfigBuilder(name="demo")
builder.l1.spacy(model="en_core_web_sm")
builder.l3.configure(model="knowledgator/gliner-linker-large-v1.0")

executor = DAGExecutor(builder.get_config())
executor.load_entities("data/entities.jsonl", target_layers=["dict"])

result = executor.execute({
    "texts": ["Farnese Palace is one of the most important palaces in the city of Rome."]
})

l0_result = result.get("l0_result")
for entity in l0_result.entities:
    if entity.linked_entity:
        print(f"{entity.mention_text} → {entity.linked_entity.label}")
        print(f"  Confidence: {entity.linked_entity.score:.3f}")

Reranking Models (L4)

When the candidate set from L2 is large, a single GLiNER call may be impractical. The L4 reranker splits candidates into chunks of max_labels and runs inference on each chunk, then merges and deduplicates results.

Model▲	Base Encoder▲	Use Case▲
gliner-linker-rerank-v1.0	ettin-encoder-68m	Reranking

Usage in Pipeline

from glinker import ProcessorFactory

executor = ProcessorFactory.create_simple(
    model_name="knowledgator/gliner-bi-base-v2.0",
    threshold=0.5,
    reranker_model="knowledgator/gliner-linker-rerank-v1.0",
    reranker_max_labels=20,
    reranker_threshold=0.3,
    entities="data/entities.jsonl",
    precompute_embeddings=True,
)

result = executor.execute({
    "texts": ["BRCA1 mutations are linked to breast cancer susceptibility."]
})

Model Selection Guide

Use Case	NER Model	Linker Model	Reranker
Edge / Mobile	gliner-bi-edge-v2.0	gliner-linker-base-v1.0	—
Balanced	gliner-bi-small-v2.0	gliner-linker-base-v1.0	—
High Accuracy	gliner-bi-base-v2.0	gliner-linker-large-v1.0	Optional
Maximum Accuracy	gliner-bi-large-v2.0	gliner-linker-large-v1.0	gliner-linker-rerank-v1.0

tip

For most use cases, the balanced configuration (gliner-bi-small-v2.0 + gliner-linker-base-v1.0) provides the best trade-off between speed and accuracy. Add the reranker only when disambiguation accuracy is critical and the candidate set is large.