normalization/sample_test_split.py

#!/usr/bin/env python3
"""
Sample 1000 samples per language from merged_transliteration_sampled.jsonl
with constraints:
- Each sample must have < 100 words (counting both input_text and output_text)
- Samples must not exist in sampled_100k_translit.jsonl (train split)
"""

import json
import random
from collections import defaultdict
from typing import Set, Dict, List, Tuple

def count_words(text: str) -> int:
    """Count words in a text string."""
    if not text:
        return 0
    return len(text.split())

def load_train_split(train_file: str) -> Set[Tuple[str, str]]:
    """
    Load train split and create a set of (input_text, output_text) tuples
    for fast lookup to avoid duplicates.
    """
    train_samples = set()
    print(f"Loading train split from {train_file}...")
    with open(train_file, 'r', encoding='utf-8') as f:
        for line_num, line in enumerate(f, 1):
            if line_num % 10000 == 0:
                print(f"  Processed {line_num} lines...")
            try:
                data = json.loads(line.strip())
                input_text = data.get('input_text', '').strip()
                output_text = data.get('output_text', '').strip()
                # Create a tuple for deduplication
                train_samples.add((input_text, output_text))
            except json.JSONDecodeError:
                continue
    print(f"Loaded {len(train_samples)} samples from train split")
    return train_samples

def sample_test_split(
    source_file: str,
    train_samples: Set[Tuple[str, str]],
    samples_per_language: int = 1000,
    max_words: int = 100
) -> Dict[str, List[Dict]]:
    """
    Sample test split from source file.
    
    Returns:
        Dictionary mapping language to list of sampled samples
    """
    # Group samples by language
    samples_by_language = defaultdict(list)
    
    print(f"\nReading source file: {source_file}")
    with open(source_file, 'r', encoding='utf-8') as f:
        for line_num, line in enumerate(f, 1):
            if line_num % 100000 == 0:
                print(f"  Processed {line_num} lines...")
            try:
                data = json.loads(line.strip())
                input_text = data.get('input_text', '').strip()
                output_text = data.get('output_text', '').strip()
                language = data.get('language', '').strip()
                
                if not language or not input_text or not output_text:
                    continue
                
                # Check word count
                total_words = count_words(input_text) + count_words(output_text)
                if total_words >= max_words:
                    continue
                
                # Check if sample exists in train split
                sample_tuple = (input_text, output_text)
                if sample_tuple in train_samples:
                    continue
                
                # Add to language group
                samples_by_language[language].append(data)
            except json.JSONDecodeError:
                continue
    
    print(f"\nFound samples by language:")
    for lang, samples in samples_by_language.items():
        print(f"  {lang}: {len(samples)} samples")
    
    # Sample from each language
    sampled_data = {}
    print(f"\nSampling {samples_per_language} samples per language...")
    for language, samples in samples_by_language.items():
        if len(samples) < samples_per_language:
            print(f"  WARNING: {language} has only {len(samples)} samples, "
                  f"requested {samples_per_language}. Using all available.")
            sampled_data[language] = samples
        else:
            sampled_data[language] = random.sample(samples, samples_per_language)
        print(f"  {language}: sampled {len(sampled_data[language])} samples")
    
    return sampled_data

def write_output(sampled_data: Dict[str, List[Dict]], output_file: str):
    """Write sampled data to output file."""
    print(f"\nWriting output to {output_file}...")
    total_samples = 0
    with open(output_file, 'w', encoding='utf-8') as f:
        for language, samples in sorted(sampled_data.items()):
            for sample in samples:
                f.write(json.dumps(sample, ensure_ascii=False) + '\n')
                total_samples += 1
    print(f"Written {total_samples} samples to {output_file}")

def main():
    source_file = "/projects/data/Embedding/IndicToolkit/datasets_final/data/merged_transliteration_sampled.jsonl"
    train_file = "/projects/data/Embedding/IndicToolkit/datasets_final/data/sampled_100k_translit.jsonl"
    output_file = "/projects/data/Embedding/IndicToolkit/datasets_final/data/test_split_translit.jsonl"
    
    # Set random seed for reproducibility
    random.seed(42)
    
    # Load train split to avoid duplicates
    train_samples = load_train_split(train_file)
    
    # Sample test split
    sampled_data = sample_test_split(
        source_file=source_file,
        train_samples=train_samples,
        samples_per_language=1000,
        max_words=100
    )
    
    # Write output
    write_output(sampled_data, output_file)
    
    print("\nDone!")

if __name__ == "__main__":
    main()