juyoung-trl/eval-dataset

from datasets import load_dataset, Dataset
from itertools import islice
from operator import itemgetter
import re

def extract_dump_date(dump_str):
    """Extract date from dump string like 'CC-MAIN-2024-10'"""
    if not dump_str:
        return '0'  # For items without dump field
    match = re.search(r'(\d{4}-\d{2})', dump_str)
    return match.group(1) if match else '0'

def load_and_verify_datasets():
    # Load streaming datasets
    datasets = {
        "ko": load_dataset("HuggingFaceFW/fineweb-2", "kor_Hang", split="test", streaming=True),
        "zh": load_dataset("HuggingFaceFW/fineweb-2", "cmn_Hani", split="test", streaming=True),
        "en": load_dataset("HuggingFaceFW/fineweb-edu", "CC-MAIN-2024-10", split="train", streaming=True),
        "ja": load_dataset("HuggingFaceFW/fineweb-2", "jpn_Jpan", split="test", streaming=True),
    }
    
    processed_datasets = {}
    
    for lang, ds in datasets.items():
        print(f"\nProcessing {lang} dataset...")
        
        # Collect items with their dumps
        items_with_dumps = []
        for item in islice(ds, 100000):  # Collect first 100K items
            dump = item.get('dump', '')
            items_with_dumps.append((item, dump))
            
        # Sort by dump date in descending order (most recent first)
        items_with_dumps.sort(key=lambda x: extract_dump_date(x[1]), reverse=True)
        
        # Print dump distribution of sorted data
        print("\nDump distribution (most recent first):")
        dump_counts = {}
        for _, dump in items_with_dumps[:1000]:  # Check first 1000 items
            dump_counts[dump] = dump_counts.get(dump, 0) + 1
        for dump, count in sorted(dump_counts.items(), key=lambda x: extract_dump_date(x[0]), reverse=True):
            print(f"  {dump}: {count} items")
        
        texts_set = set()
        filtered_texts = []
        
        # Process sorted items
        for item, dump in items_with_dumps:
            text = item.get('text', item.get('content', '')).strip()
            
            # Basic quality filters
            if text and len(text) > 50:  # Only keep non-empty texts with reasonable length
                texts_set.add(text)
                if len(texts_set) >= 15000:
                    break
        
        # Convert set to list and create dataset
        filtered_texts = list(texts_set)
        processed_datasets[lang] = Dataset.from_dict({"text": filtered_texts})
        print(f"\n{lang} dataset final size: {len(processed_datasets[lang])} examples")
        
        # Print sample texts with their dumps
        print("\nSample texts from most recent dump:")
        samples_shown = 0
        for item, dump in items_with_dumps:
            if samples_shown >= 2:
                break
            text = item.get('text', item.get('content', '')).strip()
            if text in texts_set:
                print(f"Dump: {dump}")
                print(f"Length: {len(text)}")
                print(f"Text preview: {text[:100]}...")
                print("---")
                samples_shown += 1
    
    return processed_datasets

def main():
    try:
        datasets = load_and_verify_datasets()
        print("\nDatasets processed with the following sizes:")
        for lang, ds in datasets.items():
            print(f"{lang}: {len(ds)} examples")

        # Create DatasetDict
        print("\nCreating DatasetDict...")
        dataset_dict = create_dataset_dict(datasets)

        # Upload to Hugging Face
        # Replace with your values
        REPO_NAME = "yourname/reponame"
        HF_TOKEN = "your_api_key"  # Don't share this token!

        print("\nUploading to Hugging Face Hub...")
        upload_to_huggingface(dataset_dict, REPO_NAME, HF_TOKEN)
        print(f"\nDataset uploaded successfully to {REPO_NAME}")


    except Exception as e:
        print(f"Error processing datasets: {str(e)}")

if __name__ == "__main__":
    main()