prepare_and_upload.py

"""
Script to prepare and upload Surah Al-Ikhlas Error Detection Dataset to Hugging Face.

Dataset: Audio recordings of Quran recitations with error labels
- 1506 WAV files
- Labels encoded in filename: ID{participant}V{verse}{T/F}
  - T = True (correct recitation)
  - F = False (contains error)
"""

import os
import re
import pandas as pd
from pathlib import Path
from datasets import Dataset, Audio, Features, Value, ClassLabel
import json

# Paths
DATASET_PATH = "/Users/muaz/Downloads/Surah Al-Ikhlas of the Holy Quran Error Detection Dataset/Dataset and Sounds"
EXCEL_PATH = os.path.join(DATASET_PATH, "Dataset.xlsx")
AUDIO_PATH = os.path.join(DATASET_PATH, "Sound recordings")
OUTPUT_PATH = "/Users/muaz/cursor/IkhlasDataset"

# Verse texts for Surah Al-Ikhlas
VERSE_TEXTS = {
    1: "قُلْ هُوَ اللَّهُ أَحَدٌ",
    2: "اللَّهُ الصَّمَدُ",
    3: "لَمْ يَلِدْ وَلَمْ يُولَدْ",
    4: "وَلَمْ يَكُن لَّهُ كُفُوًا أَحَدٌ"
}

def parse_filename(filename):
    """Parse filename like 'ID100V1F.wav' to extract components."""
    match = re.match(r'ID(\d+)V(\d)([TF])\.wav', filename)
    if match:
        return {
            'participant_id': int(match.group(1)),
            'verse': int(match.group(2)),
            'is_correct': match.group(3) == 'T',
            'label': 1 if match.group(3) == 'T' else 0,
            'label_text': 'correct' if match.group(3) == 'T' else 'error'
        }
    return None

def load_and_prepare_data():
    """Load audio files and Excel metadata."""
    print("=" * 60)
    print("Loading and preparing dataset...")
    print("=" * 60)
    
    # Get all audio files
    audio_files = list(Path(AUDIO_PATH).glob("*.wav"))
    print(f"\nFound {len(audio_files)} audio files")
    
    # Parse all filenames to extract labels
    data = []
    for audio_file in audio_files:
        parsed = parse_filename(audio_file.name)
        if parsed:
            parsed['filename'] = audio_file.name
            parsed['audio_path'] = str(audio_file)
            data.append(parsed)
    
    df = pd.DataFrame(data)
    print(f"Parsed {len(df)} files successfully")
    
    # Load Excel for additional metadata
    print("\nLoading Excel metadata...")
    excel_df = pd.read_excel(EXCEL_PATH, sheet_name='Sheet1')
    excel_df.columns = [col.strip() for col in excel_df.columns]
    
    # Sort both dataframes to align
    df = df.sort_values(['participant_id', 'verse']).reset_index(drop=True)
    
    # Add verse text
    df['verse_text'] = df['verse'].map(VERSE_TEXTS)
    
    # Since the Excel has 1506 rows and we have 1506 files, try direct assignment
    if len(df) == len(excel_df):
        excel_df = excel_df.reset_index(drop=True)
        
        # Add error information from Excel
        df['error_type'] = excel_df['Error type'].apply(lambda x: '' if x == 0 or pd.isna(x) else str(x))
        df['error_location'] = excel_df['Error location'].apply(lambda x: '' if x == 0 or pd.isna(x) else str(x))
        df['error_explanation'] = excel_df['Error explanation'].apply(lambda x: '' if x == 0 or pd.isna(x) else str(x))
        df['error_count'] = excel_df['Error number'].fillna(0).astype(int)
        
        print("Merged Excel metadata successfully!")
    else:
        df['error_type'] = ''
        df['error_location'] = ''
        df['error_explanation'] = ''
        df['error_count'] = 0
        print(f"Warning: Excel rows ({len(excel_df)}) don't match audio files ({len(df)})")
    
    print(f"\nLabel distribution:")
    print(df['label_text'].value_counts())
    
    print(f"\nVerse distribution:")
    print(df['verse'].value_counts().sort_index())
    
    print(f"\nUnique participants: {df['participant_id'].nunique()}")
    
    return df

def create_hf_dataset(df):
    """Create Hugging Face dataset from DataFrame."""
    print("\n" + "=" * 60)
    print("Creating Hugging Face dataset...")
    print("=" * 60)
    
    # Prepare data for dataset - audio column will contain file paths
    data = {
        'audio': df['audio_path'].tolist(),
        'label': df['label'].tolist(),
        'participant_id': df['participant_id'].tolist(),
        'verse_number': df['verse'].tolist(),
        'verse_text': df['verse_text'].tolist(),
        'error_type': df['error_type'].tolist(),
        'error_location': df['error_location'].tolist(),
        'error_explanation': df['error_explanation'].tolist(),
        'error_count': df['error_count'].tolist(),
    }
    
    # Create dataset without audio feature first
    dataset = Dataset.from_dict(data)
    
    # Cast label to ClassLabel
    dataset = dataset.cast_column('label', ClassLabel(names=['error', 'correct']))
    
    # Cast audio column (this will load audio lazily)
    dataset = dataset.cast_column('audio', Audio(sampling_rate=16000))
    
    # Create train/test split (80/20) stratified by label
    dataset = dataset.train_test_split(test_size=0.2, seed=42, stratify_by_column='label')
    
    print(f"\nDataset created:")
    print(f"  Train: {len(dataset['train'])} samples")
    print(f"  Test: {len(dataset['test'])} samples")
    
    # Show label distribution in splits
    train_labels = dataset['train']['label']
    test_labels = dataset['test']['label']
    print(f"\n  Train label distribution: error={train_labels.count(0)}, correct={train_labels.count(1)}")
    print(f"  Test label distribution: error={test_labels.count(0)}, correct={test_labels.count(1)}")
    
    return dataset

def create_dataset_card():
    """Create README.md for the dataset."""
    readme_content = """---
license: cc-by-4.0
task_categories:
  - audio-classification
language:
  - ar
tags:
  - quran
  - tajweed
  - recitation
  - error-detection
  - arabic
  - audio
  - speech
  - islam
pretty_name: Surah Al-Ikhlas Quran Recitation Error Detection Dataset
size_categories:
  - 1K<n<10K
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: label
      dtype:
        class_label:
          names:
            '0': error
            '1': correct
    - name: participant_id
      dtype: int32
    - name: verse_number
      dtype: int32
    - name: verse_text
      dtype: string
    - name: error_type
      dtype: string
    - name: error_location
      dtype: string
    - name: error_explanation
      dtype: string
    - name: error_count
      dtype: int32
  splits:
    - name: train
      num_examples: 1204
    - name: test
      num_examples: 302
---

# Surah Al-Ikhlas Quran Recitation Error Detection Dataset

## Dataset Description

This dataset contains audio recordings of Quran recitations of **Surah Al-Ikhlas** (Chapter 112 - The Sincerity) with labels indicating whether each recitation contains errors in Tajweed (Quran recitation rules).

### Dataset Summary

| Statistic | Value |
|-----------|-------|
| **Total Samples** | 1,506 |
| **Correct Recitations** | 655 (43.5%) |
| **Error Recitations** | 851 (56.5%) |
| **Unique Participants** | 384 |
| **Verses** | 4 |
| **Audio Format** | WAV |
| **Language** | Arabic |

### Surah Al-Ikhlas Text

| Verse | Arabic | Transliteration | Translation |
|-------|--------|-----------------|-------------|
| 1 | قُلْ هُوَ اللَّهُ أَحَدٌ | Qul huwa Allahu ahad | Say, "He is Allah, [who is] One" |
| 2 | اللَّهُ الصَّمَدُ | Allahu assamad | "Allah, the Eternal Refuge" |
| 3 | لَمْ يَلِدْ وَلَمْ يُولَدْ | Lam yalid walam yulad | "He neither begets nor is born" |
| 4 | وَلَمْ يَكُن لَّهُ كُفُوًا أَحَدٌ | Walam yakun lahu kufuwan ahad | "Nor is there to Him any equivalent" |

## Dataset Structure

### Data Fields

| Field | Type | Description |
|-------|------|-------------|
| `audio` | Audio | Audio file (WAV format, 16kHz) |
| `label` | ClassLabel | 0 = error, 1 = correct |
| `participant_id` | int32 | Unique identifier for the reciter |
| `verse_number` | int32 | Verse number (1-4) |
| `verse_text` | string | Arabic text of the verse |
| `error_type` | string | Type of Tajweed error (Arabic, if applicable) |
| `error_location` | string | Location of error in the verse |
| `error_explanation` | string | Explanation of the error (Arabic) |
| `error_count` | int32 | Number of errors in the recitation |

### Data Splits

| Split | Samples | Error | Correct |
|-------|---------|-------|---------|
| Train | 1,204 | 680 | 524 |
| Test | 302 | 171 | 131 |

## Usage

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("YOUR_USERNAME/surah-al-ikhlas-error-detection")

# Access training data
train_data = dataset['train']

# Example: Get first sample
sample = train_data[0]
print(f"Label: {'Correct' if sample['label'] == 1 else 'Error'}")
print(f"Verse {sample['verse_number']}: {sample['verse_text']}")
print(f"Audio sampling rate: {sample['audio']['sampling_rate']} Hz")

# Filter by label
correct_samples = train_data.filter(lambda x: x['label'] == 1)
error_samples = train_data.filter(lambda x: x['label'] == 0)
```

### Training Example

```python
from datasets import load_dataset
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification, TrainingArguments, Trainer
import evaluate

# Load dataset
dataset = load_dataset("YOUR_USERNAME/surah-al-ikhlas-error-detection")

# Load model and feature extractor
model_name = "facebook/wav2vec2-base"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
model = AutoModelForAudioClassification.from_pretrained(
    model_name,
    num_labels=2,
    label2id={"error": 0, "correct": 1},
    id2label={0: "error", 1: "correct"}
)

# Preprocess
def preprocess(examples):
    audio_arrays = [x["array"] for x in examples["audio"]]
    inputs = feature_extractor(
        audio_arrays,
        sampling_rate=16000,
        padding=True,
        return_tensors="pt"
    )
    inputs["labels"] = examples["label"]
    return inputs

dataset = dataset.map(preprocess, batched=True, remove_columns=["audio"])

# Train
training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",
    num_train_epochs=5,
    per_device_train_batch_size=8,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
)

trainer.train()
```

## Applications

This dataset can be used for:

- 🎯 Training audio classification models for Tajweed error detection
- 📱 Building Quran recitation assessment applications
- 🔬 Research in Arabic speech processing
- 📚 Educational tools for learning proper Quran recitation
- 🤖 Developing AI-assisted Quran tutoring systems

## Error Types

The dataset includes various Tajweed errors such as:
- Errors in Qalqalah (قلقلة) - echoing sounds
- Errors in letter pronunciation
- Errors in elongation (Madd)
- And other Tajweed rule violations

## Citation

If you use this dataset in your research, please cite it appropriately.

## License

This dataset is released under the [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) license.
"""
    
    readme_path = os.path.join(OUTPUT_PATH, "README.md")
    with open(readme_path, 'w', encoding='utf-8') as f:
        f.write(readme_content)
    
    print(f"\nDataset card created: {readme_path}")
    return readme_path

def main():
    # Step 1: Load and prepare data
    df = load_and_prepare_data()
    
    # Step 2: Create Hugging Face dataset
    dataset = create_hf_dataset(df)
    
    # Step 3: Create dataset card
    create_dataset_card()
    
    # Step 4: Save metadata CSV for reference
    metadata_path = os.path.join(OUTPUT_PATH, "metadata.csv")
    df.to_csv(metadata_path, index=False)
    print(f"\nMetadata saved to: {metadata_path}")
    
    # Step 5: Print upload instructions
    print("\n" + "=" * 60)
    print("READY TO UPLOAD TO HUGGING FACE")
    print("=" * 60)
    print("""
Your dataset is prepared! To upload to Hugging Face:

1. First, login to Hugging Face CLI:
   
   huggingface-cli login

2. Then run the upload script:
   
   python3 upload_to_hf.py

Make sure to edit upload_to_hf.py and replace YOUR_USERNAME with your 
actual Hugging Face username before running!
""")
    
    # Create upload script
    upload_script = '''"""Upload dataset to Hugging Face Hub"""
import os
import re
import pandas as pd
from pathlib import Path
from datasets import Dataset, Audio, ClassLabel

# Configuration - CHANGE THIS TO YOUR USERNAME
HF_USERNAME = "YOUR_USERNAME"  # <-- Change this!
REPO_NAME = "surah-al-ikhlas-error-detection"

# Paths
DATASET_PATH = "/Users/muaz/Downloads/Surah Al-Ikhlas of the Holy Quran Error Detection Dataset/Dataset and Sounds"
EXCEL_PATH = os.path.join(DATASET_PATH, "Dataset.xlsx")
AUDIO_PATH = os.path.join(DATASET_PATH, "Sound recordings")

VERSE_TEXTS = {
    1: "قُلْ هُوَ اللَّهُ أَحَدٌ",
    2: "اللَّهُ الصَّمَدُ",
    3: "لَمْ يَلِدْ وَلَمْ يُولَدْ",
    4: "وَلَمْ يَكُن لَّهُ كُفُوًا أَحَدٌ"
}

def parse_filename(filename):
    match = re.match(r'ID(\\d+)V(\\d)([TF])\\.wav', filename)
    if match:
        return {
            'participant_id': int(match.group(1)),
            'verse': int(match.group(2)),
            'label': 1 if match.group(3) == 'T' else 0,
        }
    return None

print("Loading dataset...")

# Get all audio files
audio_files = list(Path(AUDIO_PATH).glob("*.wav"))

# Parse all filenames
data = []
for audio_file in audio_files:
    parsed = parse_filename(audio_file.name)
    if parsed:
        parsed['audio_path'] = str(audio_file)
        data.append(parsed)

df = pd.DataFrame(data)
df = df.sort_values(['participant_id', 'verse']).reset_index(drop=True)
df['verse_text'] = df['verse'].map(VERSE_TEXTS)

# Load Excel metadata
excel_df = pd.read_excel(EXCEL_PATH, sheet_name='Sheet1')
excel_df.columns = [col.strip() for col in excel_df.columns]
excel_df = excel_df.reset_index(drop=True)

df['error_type'] = excel_df['Error type'].apply(lambda x: '' if x == 0 or pd.isna(x) else str(x))
df['error_location'] = excel_df['Error location'].apply(lambda x: '' if x == 0 or pd.isna(x) else str(x))
df['error_explanation'] = excel_df['Error explanation'].apply(lambda x: '' if x == 0 or pd.isna(x) else str(x))
df['error_count'] = excel_df['Error number'].fillna(0).astype(int)

print(f"Loaded {len(df)} samples")

# Create dataset
dataset_dict = {
    'audio': df['audio_path'].tolist(),
    'label': df['label'].tolist(),
    'participant_id': df['participant_id'].tolist(),
    'verse_number': df['verse'].tolist(),
    'verse_text': df['verse_text'].tolist(),
    'error_type': df['error_type'].tolist(),
    'error_location': df['error_location'].tolist(),
    'error_explanation': df['error_explanation'].tolist(),
    'error_count': df['error_count'].tolist(),
}

dataset = Dataset.from_dict(dataset_dict)
dataset = dataset.cast_column('label', ClassLabel(names=['error', 'correct']))
dataset = dataset.cast_column('audio', Audio(sampling_rate=16000))

# Split
dataset = dataset.train_test_split(test_size=0.2, seed=42, stratify_by_column='label')

print(f"Train: {len(dataset['train'])}, Test: {len(dataset['test'])}")

# Upload
print(f"\\nUploading to {HF_USERNAME}/{REPO_NAME}...")
dataset.push_to_hub(
    f"{HF_USERNAME}/{REPO_NAME}",
    private=False
)

print(f"\\n✅ Upload complete!")
print(f"View your dataset at: https://huggingface.co/datasets/{HF_USERNAME}/{REPO_NAME}")
'''
    
    upload_path = os.path.join(OUTPUT_PATH, "upload_to_hf.py")
    with open(upload_path, 'w') as f:
        f.write(upload_script)
    
    print(f"Upload script created: {upload_path}")
    
    return dataset

if __name__ == "__main__":
    dataset = main()