Training

Fine-tune a GLiFormer checkpoint on annotated data with train_model. This example adapts the model for named entity recognition.

Prepare Training Data

NER records contain a text and an extraction list. Each extraction group supplies candidate labels through all_labels and entity mention/label pairs through ner:

train_data = [
    {
        "text": "Alice works at Acme.",
        "extraction": [
            {
                "name": "entities",
                "all_labels": ["person", "organization", "location"],
                "ner": [["Alice", "person"], ["Acme", "organization"]],
            }
        ],
    },
    {
        "text": "Bob lives in Berlin.",
        "extraction": [
            {
                "name": "entities",
                "all_labels": ["person", "organization", "location"],
                "ner": [["Bob", "person"], ["Berlin", "location"]],
            }
        ],
    },
]

This small dataset illustrates the format. For your task, use a representative training set and separate held-out examples for evaluation.

Fine-Tune and Save

from gliformer import GLiFormer

model = GLiFormer.from_pretrained(
    "knowledgator/gliformer-base-v1",
    load_tokenizer=True,
)

trainer = model.train_model(
    train_dataset=train_data,
    output_dir="outputs/ner",
    max_steps=100,
    per_device_train_batch_size=2,
    learning_rate=1e-5,
)
model.save_pretrained("outputs/ner/final")

Reload the saved model with GLiFormer.from_pretrained("outputs/ner/final", load_tokenizer=True).

Training Options

Option	Purpose
eval_dataset	Held-out annotated examples for evaluation
freeze_components=["text_encoder"]	Freeze the text encoder during training
train_head_only=True	Train task heads while freezing other components
resume_from_checkpoint="path/to/training-checkpoint"	Resume a saved training run

Annotations for other tasks must match their processors and enabled heads. See the task processors, sample records, and training configurations for additional formats and settings.

Evaluation

Run task-specific evaluators from the GLiFormer repository. For example, after preparing CrossNER files:

python gliformer_eval/eval_ner.py \
  --model outputs/ner/final \
  --data data/NER \
  --datasets CrossNER_AI CrossNER_literature CrossNER_music CrossNER_politics CrossNER_science \
  --output eval_results/gliformer_ner.json

Each dataset directory must contain labels.json and test.json. See the evaluation guide for classification, relation extraction, structuring, and similarity evaluation.