Components & Configs¶

GLiNER supports multiple architecture variants, each with its own configuration class. This page documents the configuration parameters for each architecture and provides training examples.

Architecture Overview¶

Architecture	Config Class	Use Case
UniEncoderSpan	`UniEncoderSpanConfig`	Standard span-based NER, original GLiNER
UniEncoderToken	`UniEncoderTokenConfig`	Token-level NER, long-form extraction
BiEncoderSpan	`BiEncoderSpanConfig`	Span NER with separate label encoder
BiEncoderToken	`BiEncoderTokenConfig`	Token NER with separate label encoder
UniEncoderSpanDecoder	`UniEncoderSpanDecoderConfig`	Generative label prediction
UniEncoderSpanRelex	`UniEncoderSpanRelexConfig`	Joint entity and relation extraction

Base Configuration Parameters¶

All GLiNER architectures share these base configuration parameters from BaseGLiNERConfig:

Core Parameters¶

`model_name`¶

str, optional, defaults to "microsoft/deberta-v3-small"

Base encoder model identifier from Hugging Face Hub or local path.

`name`¶

str, optional, defaults to "gliner"

Optional display name for this model configuration.

`max_width`¶

int, optional, defaults to 12

Maximum span width (in number of tokens) allowed when generating candidate spans. Only applies to span-based architectures.

`hidden_size`¶

int, optional, defaults to 512

Dimensionality of hidden representations in internal layers.

`dropout`¶

float, optional, defaults to 0.4

Dropout rate applied to intermediate layers.

`fine_tune`¶

bool, optional, defaults to True

Whether to fine-tune the encoder during training.

`subtoken_pooling`¶

str, optional, defaults to "first"

Currently only first token pooling is supported. More approaches will be added in the future.

`span_mode` ^[source]¶

str, optional, defaults to "markerV0"

Defines the strategy for constructing span representations from encoder outputs. Only applies to span-based architectures.

Available options:

"markerV0" — Projects the start and end token representations with MLPs, concatenates them, and then applies a final projection. Lightweight and default.
"marker" — Similar to markerV0 but with deeper two-layer projections; better for complex tasks.
"query" — Uses learned per-span-width query vectors and dot-product interaction.
"mlp" — Applies a feedforward MLP and reshapes output into span format; fast but position-agnostic.
"cat" — Concatenates token features with learned span width embeddings before projection.
"conv_conv" — Uses multiple 1D convolutions with increasing kernel sizes; captures internal structure.
"conv_max" — Max pooling over tokens in span; emphasizes the strongest token.
"conv_mean" — Mean pooling across span tokens.
"conv_sum" — Sum pooling; raw additive representation.
"conv_share" — Shared convolution kernel over span widths; parameter-efficient alternative.

`post_fusion_schema` ^[source]¶

str, optional, defaults to ""

Defines the multi-step attention schema used to fuse span and label embeddings. The value is a string with hyphen-separated tokens that determine the sequence of attention operations applied in the CrossFuser module.

Each token in the schema defines one of the following attention types:

"l2l" — label-to-label self-attention (intra-label interaction)
"t2t" — token-to-token self-attention (intra-span interaction)
"l2t" — label-to-token cross-attention (labels attend to span tokens)
"t2l" — token-to-label cross-attention (tokens attend to labels)

Examples:

"l2l-l2t-t2t" — apply label self-attention → label-to-token attention → token self-attention
"l2t" — a single step where labels attend to span tokens
"" — disables fusion entirely (no interaction is applied)

The number of fusion layers (num_post_fusion_layers) controls how many times the entire schema is repeated.

`num_post_fusion_layers`¶

int, optional, defaults to 1

Number of layers applied after span-label fusion.

`vocab_size`¶

int, optional, defaults to -1

Vocabulary size override if needed. Automatically set during model initialization.

`max_neg_type_ratio`¶

int, optional, defaults to 1

Controls the ratio of negative (non-matching) types during training.

`max_types`¶

int, optional, defaults to 25

Maximum number of entity types supported per batch.

`max_len`¶

int, optional, defaults to 384

Maximum sequence length accepted by the encoder.

`words_splitter_type`¶

str, optional, defaults to "whitespace"

Heuristic used for word-level splitting during inference.
Choices: "whitespace", "spacy", "moses", stanza, universal

`num_rnn_layers`¶

int, optional, defaults to 1

Number of LSTM layers to apply on top of encoder outputs. Set to 0 to disable LSTM.

`fuse_layers`¶

bool, optional, defaults to False

If True, combine representations from multiple encoders (labels and main encoder).

`embed_ent_token`¶

bool, optional, defaults to True

If True, <<ENT>> tokens will be pooled for each label. If False, the first token of each label will be pooled as label embedding.

`class_token_index`¶

int, optional, defaults to -1

Index of the entity token in the vocabulary. Set automatically during initialization.

`encoder_config`¶

dict or PretrainedConfig, optional

A nested config dictionary for the encoder model. If a dict is passed, its model_type must be set or inferred.

`ent_token`¶

str, optional, defaults to "<<ENT>>"

Special token used to mark entity type boundaries in the input.

`sep_token`¶

str, optional, defaults to "<<SEP>>"

Token used to separate entity types from input text.

`_attn_implementation`¶

str, optional

Optional override for attention logic. Can be used to disable Flash Attention if installed.

Example:

model = GLiNER.from_pretrained(
    "urchade/gliner_mediumv2.1", 
    _attn_implementation="eager"  # Disable Flash Attention
)

UniEncoder Span Configuration¶

UniEncoderSpanConfig is used for the original GLiNER architecture with span-based prediction.

Architecture-Specific Parameters¶

This architecture uses all base parameters without additional architecture-specific parameters.

Usage Example¶

from gliner import GLiNERConfig, GLiNER

# Create config for UniEncoderSpan
config = GLiNERConfig(
    model_name="microsoft/deberta-v3-small",
    max_width=12,
    hidden_size=512,
    span_mode="markerV0",
    # labels_encoder=None  # Makes it UniEncoder
    # labels_decoder=None  # No decoder
    # relations_layer=None  # No relations
)

# Initialize model from config
model = GLiNER.from_config(config)

Training Config Example¶

# Model Configuration
model_name: microsoft/deberta-v3-base
labels_encoder: null  # UniEncoder
name: "span level gliner"
max_width: 12
hidden_size: 768
dropout: 0.4
fine_tune: true
subtoken_pooling: first
span_mode: markerV0
post_fusion_schema: ""
num_post_fusion_layers: 1

# Training Parameters
num_steps: 30000
train_batch_size: 8
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# Loss Configuration
loss_alpha: -1
loss_gamma: 0
label_smoothing: 0
loss_reduction: "sum"

# Learning Rate Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01
max_grad_norm: 1.0

# Data Configuration
train_data: "data.json"
prev_path: null  # Training from scratch
save_total_limit: 3

# Advanced Settings
max_types: 25
max_len: 384

UniEncoder Token Configuration¶

UniEncoderTokenConfig is used for token-level classification, suitable for long-form entity extraction.

Architecture-Specific Parameters¶

`span_mode`¶

str, required, fixed to "token-level"

This parameter is automatically set to "token-level" and cannot be changed for this architecture.

Usage Example¶

from gliner import GLiNERConfig, GLiNER

# Create config for UniEncoderToken
config = GLiNERConfig(
    model_name="microsoft/deberta-v3-small",
    hidden_size=512,
    span_mode="token-level",  # Automatically set for this architecture
)

model = GLiNER.from_config(config)

Training Config Example¶

# Model Configuration
model_name: microsoft/deberta-v3-base
labels_encoder: null
name: "token level gliner"
hidden_size: 768
dropout: 0.4
fine_tune: true
subtoken_pooling: first
span_mode: token-level  # Token-level prediction
num_rnn_layers: 1  # LSTM helps with token sequences

# Training Parameters (same as span)
num_steps: 30000
train_batch_size: 8
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# Loss Configuration
loss_alpha: -1
loss_gamma: 0
label_smoothing: 0
loss_reduction: "sum"

# Learning Rate Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01
max_grad_norm: 1.0

# Data Configuration
train_data: "data.json"
prev_path: null
save_total_limit: 3

# Advanced Settings
max_types: 25
max_len: 384

BiEncoder Span Configuration¶

BiEncoderSpanConfig uses separate encoders for text and entity labels, enabling pre-computation of label embeddings.

Architecture-Specific Parameters¶

`labels_encoder`¶

str, required

Model identifier or path for the label encoder. Typically a sentence transformer model.

Examples:

"sentence-transformers/all-MiniLM-L6-v2"
"BAAI/bge-small-en-v1.5"

`labels_encoder_config`¶

dict or PretrainedConfig, optional

Nested configuration for the label encoder model.

Important Notes¶

Unlike UniEncoder models, BiEncoder models do not support token embedding resizing. The vocabulary is fixed to the pretrained encoder’s vocabulary.

Usage Example¶

from gliner import GLiNERConfig, GLiNER

# Create config for BiEncoderSpan
config = GLiNERConfig(
    model_name="microsoft/deberta-v3-base",
    labels_encoder="sentence-transformers/all-MiniLM-L6-v2",  # Bi-encoder
    max_width=12,
    hidden_size=768,
    span_mode="markerV0",
)

model = GLiNER.from_config(config)

# Pre-compute label embeddings for efficiency
labels = ["person", "organization", "location"]
labels_embeddings = model.encode_labels(labels)

# Use pre-computed embeddings for inference
entities = model.batch_predict_with_embeds(
    texts=["Apple Inc. was founded by Steve Jobs."],
    labels_embeddings=labels_embeddings,
    labels=labels
)

Training Config Example¶

# Model Configuration
model_name: microsoft/deberta-v3-base
labels_encoder: sentence-transformers/all-MiniLM-L6-v2  # Bi-encoder
name: "bi-encoder span gliner"
max_width: 12
hidden_size: 768
dropout: 0.4
fine_tune: true
subtoken_pooling: first
span_mode: markerV0
post_fusion_schema: "l2t-t2l"  # Cross-attention fusion

# Training Parameters
num_steps: 30000
train_batch_size: 8
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# Loss Configuration (Focal loss recommended)
loss_alpha: 0.25
loss_gamma: 2.0
label_smoothing: 0
loss_reduction: "sum"

# Learning Rate Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01
max_grad_norm: 1.0

# Data Configuration
train_data: "data.json"
prev_path: null
save_total_limit: 3

# Advanced Settings
max_types: 100  # Can handle many more types
max_len: 384

BiEncoder Token Configuration¶

BiEncoderTokenConfig combines bi-encoder architecture with token-level prediction.

Architecture-Specific Parameters¶

`labels_encoder`¶

str, required

Model identifier for the label encoder.

`span_mode`¶

str, required, fixed to "token-level"

Automatically set to "token-level" for this architecture.

Usage Example¶

from gliner import GLiNERConfig, GLiNER

# Create config for BiEncoderToken
config = GLiNERConfig(
    model_name="microsoft/deberta-v3-base",
    labels_encoder="sentence-transformers/all-MiniLM-L6-v2",
    hidden_size=768,
    span_mode="token-level",
)

model = GLiNER.from_config(config)

Training Config Example¶

# Model Configuration
model_name: microsoft/deberta-v3-base
labels_encoder: sentence-transformers/all-MiniLM-L6-v2
name: "bi-encoder token gliner"
hidden_size: 768
dropout: 0.4
fine_tune: true
subtoken_pooling: first
span_mode: token-level
num_rnn_layers: 1

# Training Parameters
num_steps: 30000
train_batch_size: 8
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# Loss Configuration
loss_alpha: 0.25
loss_gamma: 2.0
label_smoothing: 0
loss_reduction: "sum"

# Learning Rate Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01
max_grad_norm: 1.0

# Data Configuration
train_data: "data.json"
prev_path: null
save_total_limit: 3

# Advanced Settings
max_types: 100
max_len: 384

UniEncoder Span Decoder Configuration¶

UniEncoderSpanDecoderConfig extends span-based NER with a generative decoder for label generation.

Architecture-Specific Parameters¶

`labels_decoder`¶

str, required

Model identifier for the generative decoder (e.g., GPT-2).

Examples:

"gpt2"
"distilgpt2"
"EleutherAI/gpt-neo-125M"

`decoder_mode`¶

str, optional

Defines how decoder inputs are constructed.

Choices:

"prompt" — Use entity type embeddings as decoder context
"span" — Use span token representations as decoder context

`full_decoder_context`¶

bool, optional, defaults to True

Whether to provide full context to the decoder (all tokens in span) or just boundary markers.

`blank_entity_prob`¶

float, optional, defaults to 0.1

Probability of using a generic “entity” label during training for improved generalization.

`labels_decoder_config`¶

dict or PretrainedConfig, optional

Nested configuration for the decoder model.

`decoder_loss_coef`¶

float, optional, defaults to 0.5

Weight for the decoder generation loss in the total loss.

`span_loss_coef`¶

float, optional, defaults to 0.5

Weight for the span classification loss in the total loss.

Usage Example¶

from gliner import GLiNERConfig, GLiNER

# Create config for UniEncoderSpanDecoder
config = GLiNERConfig(
    model_name="microsoft/deberta-v3-base",
    labels_decoder="gpt2",  # Add decoder
    decoder_mode="span",
    full_decoder_context=True,
    blank_entity_prob=0.1,
    decoder_loss_coef=0.5,
    span_loss_coef=0.5,
)

model = GLiNER.from_config(config)

Training Config Example¶

# Model Configuration
model_name: microsoft/deberta-v3-base
labels_decoder: gpt2  # Generative decoder
decoder_mode: span
full_decoder_context: true
blank_entity_prob: 0.1
name: "span decoder gliner"
max_width: 12
hidden_size: 768
dropout: 0.4
fine_tune: true
span_mode: markerV0

# Loss Configuration
decoder_loss_coef: 0.5
span_loss_coef: 0.5

# Training Parameters
num_steps: 30000
train_batch_size: 4  # Smaller due to decoder
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# Loss Configuration
loss_alpha: -1
loss_gamma: 0
label_smoothing: 0.1  # Helps with generation
loss_reduction: "sum"

# Learning Rate Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01
max_grad_norm: 1.0

# Data Configuration
train_data: "data.json"
prev_path: null
save_total_limit: 3

# Advanced Settings
max_types: 25
max_len: 384

UniEncoder Span Relex Configuration¶

UniEncoderSpanRelexConfig extends span-based NER with relation extraction capabilities.

Architecture-Specific Parameters¶

`relations_layer`¶

str, required

Type of relation representation layer to use.

Choices:

"dot" — Dot product between entity representations
"gcn" — Graph convolutional network for modeling interactions between entities
"gat" — Graph attention network for modeling interactions between entities

`triples_layer`¶

str, optional

Type of triple scoring layer for (head, relation, tail) scoring.

Choices:

"distmult" — DistMult scoring function
"complex" — ComplEx scoring function
"transe" — TransE scoring function

`embed_rel_token`¶

bool, optional, defaults to True

Whether to embed relation type tokens similar to entity tokens.

`rel_token_index`¶

int, optional, defaults to -1

Index of the relation token in vocabulary. Set automatically during initialization.

`rel_token`¶

str, optional, defaults to "<<REL>>"

Special token used to mark relation types in the input.

`span_loss_coef`¶

float, optional, defaults to 1.0

Weight for entity span classification loss.

`adjacency_loss_coef`¶

float, optional, defaults to 1.0

Weight for entity pair adjacency prediction loss.

`relation_loss_coef`¶

float, optional, defaults to 1.0

Weight for relation type classification loss.

Usage Example¶

from gliner import GLiNERConfig, GLiNER

# Create config for UniEncoderSpanRelex
config = GLiNERConfig(
    model_name="microsoft/deberta-v3-base",
    relations_layer="biaffine",  # Enable relations
    triples_layer="distmult",
    rel_token="<<REL>>",
    span_loss_coef=1.0,
    adjacency_loss_coef=1.0,
    relation_loss_coef=1.0,
)

model = GLiNER.from_config(config)

Training Config Example¶

# Model Configuration
model_name: microsoft/deberta-v3-base
relations_layer: biaffine  # Enable relation extraction
triples_layer: distmult
rel_token: "<<REL>>"
embed_rel_token: true
name: "span relex gliner"
max_width: 12
hidden_size: 768
dropout: 0.4
fine_tune: true
span_mode: markerV0

# Loss Configuration
span_loss_coef: 1.0
adjacency_loss_coef: 1.0
relation_loss_coef: 1.0

# Training Parameters
num_steps: 30000
train_batch_size: 6  # Smaller due to relation computation
eval_every: 1000
warmup_ratio: 0.1
scheduler_type: "cosine"

# Loss Configuration
loss_alpha: -1
loss_gamma: 0
label_smoothing: 0
loss_reduction: "sum"

# Learning Rate Configuration
lr_encoder: 1e-5
lr_others: 5e-5
weight_decay_encoder: 0.01
weight_decay_other: 0.01
max_grad_norm: 1.0

# Data Configuration
train_data: "data_with_relations.json"  # Must include relation annotations
prev_path: null
save_total_limit: 3

# Advanced Settings
max_types: 25
max_len: 384

Data Format for Relation Extraction¶

train_data = [
    {
        "tokenized_text": ["John", "works", "at", "Microsoft"],
        "ner": [[0, 0, "person"], [3, 3, "organization"]],
        "relations": [[0, 1, "works_at"]]  # (head_entity_idx, tail_entity_idx, relation_type)
    }
]

TrainingArguments¶

Custom extension of transformers.TrainingArguments with additional parameters for GLiNER models.

GLiNER-Specific Parameters¶

`others_lr`¶

float, optional
Learning rate for non-encoder parameters (e.g., span layers, label encoder). If not specified, uses main learning_rate.

`others_weight_decay`¶

float, optional, defaults to 0.0
Weight decay for non-encoder parameters.

`focal_loss_alpha`¶

float, optional, defaults to -1
Alpha parameter for focal loss. If ≥ 0, focal loss is activated.

Focal loss formula:
FL(p_t) = -α × (1 - p_t)^γ × log(p_t)

`focal_loss_gamma`¶

float, optional, defaults to 0
Gamma parameter for focal loss. Higher values increase focus on hard examples.

`focal_loss_prob_margin`¶

float, optional, defaults to 0.0
Probability margin for focal loss adjustment.

`label_smoothing`¶

float, optional, defaults to 0.0
Label smoothing factor ε for regularization.

`loss_reduction`¶

str, optional, defaults to "sum"
How to aggregate loss across samples.
Choices: "sum", "mean"

`negatives`¶

float, optional, defaults to 1.0
Ratio of negative to positive spans during training.

`masking`¶

str, optional, defaults to "none"
Masking strategy for negative sampling.
Choices: "none", "global", "label", "span"

Components & Configs¶

Architecture Overview¶

Base Configuration Parameters¶

Core Parameters¶

model_name¶

name¶

max_width¶

hidden_size¶

dropout¶

fine_tune¶

subtoken_pooling¶

span_mode [source]¶

post_fusion_schema [source]¶

num_post_fusion_layers¶

vocab_size¶

max_neg_type_ratio¶

max_types¶

max_len¶

words_splitter_type¶

num_rnn_layers¶

fuse_layers¶

embed_ent_token¶

class_token_index¶

encoder_config¶

ent_token¶

sep_token¶

_attn_implementation¶

UniEncoder Span Configuration¶

Architecture-Specific Parameters¶

Usage Example¶

Training Config Example¶

UniEncoder Token Configuration¶

Architecture-Specific Parameters¶

span_mode¶

Usage Example¶

Training Config Example¶

BiEncoder Span Configuration¶

Architecture-Specific Parameters¶

labels_encoder¶

labels_encoder_config¶

Important Notes¶

Usage Example¶

Training Config Example¶

BiEncoder Token Configuration¶

Architecture-Specific Parameters¶

labels_encoder¶

span_mode¶

Usage Example¶

Training Config Example¶

UniEncoder Span Decoder Configuration¶

Architecture-Specific Parameters¶

labels_decoder¶

decoder_mode¶

full_decoder_context¶

blank_entity_prob¶

labels_decoder_config¶

decoder_loss_coef¶

span_loss_coef¶

Usage Example¶

Training Config Example¶

UniEncoder Span Relex Configuration¶

Architecture-Specific Parameters¶

relations_layer¶

triples_layer¶

embed_rel_token¶

rel_token_index¶

rel_token¶

span_loss_coef¶

adjacency_loss_coef¶

relation_loss_coef¶

Usage Example¶

Training Config Example¶

Data Format for Relation Extraction¶

TrainingArguments¶

GLiNER-Specific Parameters¶

others_lr¶

others_weight_decay¶

focal_loss_alpha¶

focal_loss_gamma¶

focal_loss_prob_margin¶

`model_name`¶

`name`¶

`max_width`¶

`hidden_size`¶

`dropout`¶

`fine_tune`¶

`subtoken_pooling`¶

`span_mode` ^[source]¶

`post_fusion_schema` ^[source]¶

`num_post_fusion_layers`¶

`vocab_size`¶

`max_neg_type_ratio`¶

`max_types`¶

`max_len`¶

`words_splitter_type`¶

`num_rnn_layers`¶

`fuse_layers`¶

`embed_ent_token`¶

`class_token_index`¶

`encoder_config`¶

`ent_token`¶

`sep_token`¶

`_attn_implementation`¶

`span_mode`¶

`labels_encoder`¶

`labels_encoder_config`¶

`labels_encoder`¶

`span_mode`¶

`labels_decoder`¶

`decoder_mode`¶

`full_decoder_context`¶

`blank_entity_prob`¶

`labels_decoder_config`¶

`decoder_loss_coef`¶

`span_loss_coef`¶

`relations_layer`¶

`triples_layer`¶

`embed_rel_token`¶

`rel_token_index`¶

`rel_token`¶

`span_loss_coef`¶

`adjacency_loss_coef`¶

`relation_loss_coef`¶

`others_lr`¶

`others_weight_decay`¶

`focal_loss_alpha`¶

`focal_loss_gamma`¶

`focal_loss_prob_margin`¶

`label_smoothing`¶

`loss_reduction`¶

`negatives`¶

`masking`¶