csv_to_jsonl.py

#!/usr/bin/env python3
"""
Regenerate JSONL splits from dataset.csv
Useful if you lost the original JSONL files
"""

import pandas as pd
import json

def csv_to_jsonl(csv_path="dataset.csv", output_dir="."):
    df = pd.read_csv(csv_path)

    # Convert structure
    records = []
    for _, row in df.iterrows():
        record = {
            "id": row["id"],
            "translation": {
                "en": row["english"],
                "kab": row["kabyle"]
            },
            "category": row["category"],
            "subcategory": row["subcategory"],
            "domain": "road_traffic",
            "complexity": row["complexity"],
            "tokens_en": int(row["tokens_en"]),
            "tokens_kab": int(row["tokens_kab"])
        }
        records.append(record)

    # Stratified split logic (same as original)
    train, val, test = [], [], []

    for r in records:
        idx = int(r["id"].split("_")[-1])
        cat = r["category"]

        # Dangers: 70/15/15
        if cat == "dangers":
            if idx % 10 < 7:
                train.append(r)
            elif idx % 10 < 8:
                val.append(r)
            else:
                test.append(r)
        # Prohibitions: 60/20/20
        elif cat == "prohibitions":
            if idx % 10 < 6:
                train.append(r)
            elif idx % 10 < 8:
                val.append(r)
            else:
                test.append(r)
        # Obligations: 50/25/25
        elif cat == "obligations":
            if idx % 10 < 5:
                train.append(r)
            elif idx % 10 < 7:
                val.append(r)
            else:
                test.append(r)
        # End of restrictions: 40/30/30
        else:
            if idx % 10 < 4:
                train.append(r)
            elif idx % 10 < 7:
                val.append(r)
            else:
                test.append(r)

    # Save JSONL files
    for name, data in [("train.jsonl", train), ("validation.jsonl", val), ("test.jsonl", test)]:
        filepath = f"{output_dir}/{name}"
        with open(filepath, 'w', encoding='utf-8') as f:
            for item in data:
                f.write(json.dumps(item, ensure_ascii=False) + '\n')
        print(f"✓ {name}: {len(data)} entries")

    return train, val, test

if __name__ == "__main__":
    print("Regeneration des fichiers JSONL depuis dataset.csv...")
    csv_to_jsonl()
    print("\nDone!")