train.py

from pathlib import Path
import random
import shutil
from datasets import load_dataset
from transformers import TrainingArguments
from span_marker import SpanMarkerModel, Trainer
from span_marker.model_card import SpanMarkerModelCardData
from huggingface_hub import upload_folder, upload_file



def main() -> None:
    # Load the dataset, ensure "tokens" and "ner_tags" columns, and get a list of labels
    labels = ["O", "B-ORG", "I-ORG"]
    dataset_id = "tomaarsen/ner-orgs"
    dataset = load_dataset(dataset_id)

    train_dataset = dataset["train"]
    eval_dataset = dataset["validation"]
    eval_dataset = eval_dataset.select(random.sample(range(len(eval_dataset)), k=3000))
    test_dataset = dataset["test"]

    # Initialize a SpanMarker model using a pretrained BERT-style encoder
    encoder_id = "BAAI/bge-base-en-v1.5"
    model_id = "nbroad/span-marker-bge-base-orgs-v1"
    model = SpanMarkerModel.from_pretrained(
        encoder_id,
        labels=labels,
        # SpanMarker hyperparameters:
        model_max_length=256,
        marker_max_length=128,
        entity_max_length=8,
        # Model card variables
        model_card_data=SpanMarkerModelCardData(
            model_id=model_id,
            dataset_id=dataset_id,
            encoder_id=encoder_id,
            dataset_name="FewNERD, CoNLL2003, and OntoNotes v5",
            license="cc-by-sa-4.0",
            language=["en"],
        ),
    )

    # Prepare the 🤗 transformers training arguments
    output_dir = Path("models") / model_id
    args = TrainingArguments(
        output_dir=output_dir,
        run_name=model_id,
        # Training Hyperparameters:
        learning_rate=5e-5,
        per_device_train_batch_size=128,
        per_device_eval_batch_size=128,
        num_train_epochs=3,
        weight_decay=0.01,
        warmup_ratio=0.05,
        fp16=True,  # Replace `bf16` with `fp16` if your hardware can't use bf16.
        # Other Training parameters
        logging_first_step=True,
        logging_steps=100,
        evaluation_strategy="steps",
        save_strategy="steps",
        eval_steps=600,
        save_total_limit=1,
        dataloader_num_workers=4,
        metric_for_best_model="overall_f1",
        greater_is_better=True,
    )

    # Initialize the trainer using our model, training args & dataset, and train
    trainer = Trainer(
        model=model,
        args=args,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
    )
    trainer.train()

    # Compute & save the metrics on the test set
    metrics = trainer.evaluate(test_dataset, metric_key_prefix="test")
    trainer.save_metrics("test", metrics)

    # Save the model & training script locally
    trainer.save_model(output_dir / "checkpoint-final")
    shutil.copy2(__file__, output_dir / "checkpoint-final" / "train.py")

    # Upload everything to the Hub
    breakpoint()
    model.push_to_hub(model_id, private=True)
    upload_folder(folder_path=output_dir / "runs", path_in_repo="runs", repo_id=model_id)
    upload_file(path_or_fileobj=__file__, path_in_repo="train.py", repo_id=model_id)
    upload_file(path_or_fileobj=output_dir / "all_results.json", path_in_repo="all_results.json", repo_id=model_id)
    upload_file(path_or_fileobj=output_dir / "emissions.csv", path_in_repo="emissions.csv", repo_id=model_id)


if __name__ == "__main__":
    main()