prepare.py

import requests

from datasets import Dataset, DatasetDict


def load_bio_file(filepath_or_url):
    """Load a .bio file and yield structured token-label sequences."""
    # Handle both URLs and local paths
    if filepath_or_url.startswith("http"):
        response = requests.get(filepath_or_url)
        response.raise_for_status()
        lines = response.text.splitlines()
    else:
        with open(filepath_or_url, encoding="utf-8") as f:
            lines = f.readlines()

    guid = 0
    tokens, ner_tags = [], []

    for line in lines:
        line = line.strip()
        if not line or line.startswith("-DOCSTART-"):
            # End of one sentence
            if tokens:
                yield {
                    "tokens": tokens,
                    "ner_tags": ner_tags,
                }
                guid += 1
                tokens, ner_tags = [], []
        else:
            # Handle standard CoNLL-style 4-column lines
            splits = line.split()
            if len(splits) >= 2:
                token, ner = splits[:2]
                tokens.append(token)
                ner_tags.append(ner)
            else:
                # In case of malformed line
                print(f"Warning: malformed line skipped -> {line}")

    # Add last sentence if file didn't end with newline
    if tokens:
        yield {
            "tokens": tokens,
            "ner_tags": ner_tags,
        }


def load_all_splits():
    """Load train, validation, and test splits."""
    urls = {
        "train": "https://raw.githubusercontent.com/brickee/HarveyNER/refs/heads/main/data/tweets/tweets.train.bio",
        "validation": "https://raw.githubusercontent.com/brickee/HarveyNER/refs/heads/main/data/tweets/tweets.dev.bio",
        "test": "https://raw.githubusercontent.com/brickee/HarveyNER/refs/heads/main/data/tweets/tweets.test.bio",
    }

    datasets = {}
    for split, url in urls.items():
        print(f"Loading {split} data from {url} ...")
        data = list(load_bio_file(url))
        datasets[split] = data
        print(f"  Loaded {len(data)} sentences for {split}.")

    return datasets


def main():
    datasets = load_all_splits()
    print("\nExample from train:")
    print(datasets["train"][0])

    hf_datasets = DatasetDict({
        'train': Dataset.from_list(datasets['train']), 
        'validation': Dataset.from_list(datasets['validation']), 
        'test': Dataset.from_list(datasets['test']), 
    })
    hf_datasets.push_to_hub('extraordinarylab/harvey-ner')


if __name__ == "__main__":
    main()