readme.md

tar

cd /home/vladimir_albrekht/projects/2025_sep_22_qwen3omni/ms_swift_training/approach_2_transformers_based/data/fleurs_data
tar -czf - fleurs_cleaned_kk_not_translated | split -b 4G - fleurs_cleaned_kk_not_translated.tar.gz.part

cat

cat fleurs_cleaned_kk_not_translated.tar.gz.part* | tar -xzf -

processing logic

# 1. data part
import pandas as pd
from pathlib import Path

def load_fleurs_split(split_name, base_path="/home/vladimir_albrekht/projects/2025_sep_22_qwen3omni/ms_swift_training/approach_2_transformers_based/data/fleurs_data/fleurs/data/kk_kz"):
    df = pd.read_csv(
        f"{base_path}/{split_name}.tsv",
        sep="\t",
        header=None,
        names=["id", "file_name", "raw_transcription", "transcription", "phonemes", "num_samples", "gender"]
    )
    df['num_samples'] = pd.to_numeric(df['num_samples'], errors='coerce')
    df['duration_seconds'] = df['num_samples'] / 16000
    df['dataset_type'] = split_name
    df['audio_dir'] = f"{base_path}/audio/{split_name}"
    return df

df_test = load_fleurs_split("test")
df_dev = load_fleurs_split("dev")
df_train = load_fleurs_split("train")

df = pd.concat([df_test, df_dev, df_train], ignore_index=True)
df_cleaned = df[df['duration_seconds'] <= 29.9]

duplicate_files = df_cleaned['file_name'].duplicated().sum()
unique_files = df_cleaned['file_name'].nunique()
total_files = len(df_cleaned)
print(f"Total samples: {total_files}")
df_cleaned = df_cleaned.drop(columns=['phonemes', 'num_samples'])
df_cleaned = df_cleaned.rename(columns={
    'raw_transcription': 'transcription',
    'transcription': 'raw_transcription'
})


# 2. convert to .jsonl format
import shutil
from pathlib import Path

output_dir = Path("/home/vladimir_albrekht/projects/2025_sep_22_qwen3omni/ms_swift_training/approach_2_transformers_based/data/fleurs_data/fleurs_cleaned_kk")
output_dir.mkdir(exist_ok=True)
audio_output_dir = output_dir / "audios"
audio_output_dir.mkdir(exist_ok=True)

jsonl_data = []
for idx, row in df_cleaned.iterrows():
    src_audio = Path(row['audio_dir']) / row['file_name']
    new_audio_name = f"{row['dataset_type']}_{row['file_name']}"
    dst_audio = audio_output_dir / new_audio_name
    
    shutil.copy(src_audio, dst_audio)
    
    jsonl_data.append({
        "transcription": row['transcription'],
        "audio_path": f"audios/{new_audio_name}",
        "meta_data": {
            "id": row['id'],
            "raw_transcription": row['raw_transcription'],
            "duration_seconds": row['duration_seconds'],
            "gender": row['gender'],
            "dataset_type": row['dataset_type']
        }
    })

import json
with open(output_dir / "data.jsonl", 'w', encoding='utf-8') as f:
    for item in jsonl_data:
        f.write(json.dumps(item, ensure_ascii=False) + '\n')

print(f"Saved {len(jsonl_data)} samples to {output_dir}")