# TurMix & HinMix: Exact Reproduction Guide

This document provides a complete guide to reproducing the Turkish (TurMix) and Hindi (HinMix) pretraining data pipelines.

## Table of Contents
1. [Overview](#overview)
2. [Environment Setup](#environment-setup)
3. [Project Structure](#project-structure)
4. [Data Sources](#data-sources)
5. [Pipeline Stages](#pipeline-stages)
6. [Stage 1: Download](#stage-1-download)
7. [Stage 2: Quality Filtering](#stage-2-quality-filtering)
8. [Stage 3: MinHash Deduplication](#stage-3-minhash-deduplication)
9. [Stage 4: Consensus Subset Construction](#stage-4-consensus-subset-construction)
10. [Stage 5: Upload to HuggingFace](#stage-5-upload-to-huggingface)
11. [Final Statistics](#final-statistics)

---

## Overview

The pipeline processes web crawl data through the following stages:
1. **Download**: Fetch data from HuggingFace using `huggingface-cli`
2. **Quality Filter**: Apply language-specific quality filters using DataTrove
3. **MinHash Dedup**: Remove near-duplicate documents within each source
4. **Consensus**: Identify documents appearing in 2+ sources (exact text match)

### Key Design Decisions
- Use `huggingface-cli download` with `--include` patterns for efficient selective downloads
- Process each source separately to avoid parquet schema conflicts
- Use 54 workers (CPU cores - 2) for parallel processing
- MinHash deduplication within each source (not cross-source)
- Consensus detection via exact text hash matching across sources

---

## Environment Setup

### Prerequisites
```bash
# Python 3.10+
conda create -n pretraining python=3.10
conda activate pretraining

# Install dependencies
pip install datasets datatrove pyarrow huggingface_hub
pip install fasttext-langdetect  # For language detection
```

### HuggingFace Authentication
```bash
huggingface-cli login
# Enter your HuggingFace token
```

---

## Project Structure

```
arabic-pretraining-mix-other-languages/
├── run_pipeline.py              # Main unified pipeline runner
├── build_consensus_v2.py        # Consensus subset builder (memory-efficient)
├── filter_c4.py                 # C4 JSON filtering script
├── fix_hplt2_filter.py          # Turkish HPLT2 filtering (5 subfolders)
├── fix_hindi_hplt2_filter.py    # Hindi HPLT2 filtering
├── src/
│   ├── config/
│   │   ├── common.py            # Shared configs (paths, workers, MinHash params)
│   │   ├── datasets_tr.py       # Turkish dataset definitions
│   │   └── datasets_hi.py       # Hindi dataset definitions
│   ├── filters/
│   │   ├── base_quality.py      # Base quality filter class
│   │   ├── tr_quality.py        # Turkish quality filter
│   │   ├── hi_quality.py        # Hindi quality filter
│   │   └── lang_config.py       # Language-specific constants
│   └── dedup/
│       └── __init__.py          # MinHash deduplication wrappers
└── data/
    ├── hi/                      # Hindi data
    │   ├── downloads/           # Raw downloaded data
    │   ├── filtered/            # Quality-filtered data
    │   ├── deduped/             # MinHash-deduplicated data
    │   ├── consensus/           # Consensus subset
    │   └── minhash_signatures/  # MinHash signatures (preserved)
    └── tr/                      # Turkish data (same structure)
```

---

## Data Sources

### Hindi Sources (6)
| Source | HuggingFace Path | Subset/Config | Download Command |
|--------|------------------|---------------|------------------|
| HPLT-2 | `HPLT/HPLT2.0_cleaned` | `hin_Deva` | `huggingface-cli download HPLT/HPLT2.0_cleaned --include "hin_Deva/*" --local-dir ./data/hi/hplt2 --repo-type dataset` |
| Fineweb-2 | `HuggingFaceFW/fineweb-2` | `hin_Deva` | `huggingface-cli download HuggingFaceFW/fineweb-2 --include "data/hin_Deva/*" --local-dir ./data/hi/fineweb2 --repo-type dataset` |
| CulturaX | `uonlp/CulturaX` | `hi` | `huggingface-cli download uonlp/CulturaX --include "hi/*" --local-dir ./data/hi/culturax --repo-type dataset` |
| mC4 | `allenai/c4` | `hi` | `huggingface-cli download allenai/c4 --include "multilingual/c4-hi*" --local-dir ./data/hi/c4 --repo-type dataset` |
| Sangraha (verified) | `ai4bharat/sangraha` | `verified/hin` | `huggingface-cli download ai4bharat/sangraha --include "verified/hin/*" --local-dir ./data/hi/sangraha_verified --repo-type dataset` |
| Sangraha (unverified) | `ai4bharat/sangraha` | `unverified/hin` | `huggingface-cli download ai4bharat/sangraha --include "unverified/hin/*" --local-dir ./data/hi/sangraha_unverified --repo-type dataset` |

### Turkish Sources (5)
| Source | HuggingFace Path | Subset/Config | Download Command |
|--------|------------------|---------------|------------------|
| HPLT-2 | `HPLT/HPLT2.0_cleaned` | `tur_Latn` | `huggingface-cli download HPLT/HPLT2.0_cleaned --include "tur_Latn*/*" --local-dir ./data/tr/hplt2 --repo-type dataset` |
| Fineweb-2 | `HuggingFaceFW/fineweb-2` | `tur_Latn` | `huggingface-cli download HuggingFaceFW/fineweb-2 --include "data/tur_Latn/*" --local-dir ./data/tr/fineweb2 --repo-type dataset` |
| CulturaX | `uonlp/CulturaX` | `tr` | `huggingface-cli download uonlp/CulturaX --include "tr/*" --local-dir ./data/tr/culturax --repo-type dataset` |
| mC4 | `allenai/c4` | `tr` | `huggingface-cli download allenai/c4 --include "multilingual/c4-tr*" --local-dir ./data/tr/c4 --repo-type dataset` |
| VNGRS | `vngrs-ai/vngrs-web-corpus` | N/A | `huggingface-cli download vngrs-ai/vngrs-web-corpus --local-dir ./data/tr/vngrs --repo-type dataset` |

**Note**: Turkish HPLT-2 is split into 5 subfolders: `tur_Latn_1` through `tur_Latn_5`.

---

## Pipeline Stages

### Stage 1: Download

Downloads are performed using `huggingface-cli download` with `--include` patterns:

```bash
# Example: Download Hindi CulturaX
huggingface-cli download uonlp/CulturaX \
    --include "hi/*" \
    --local-dir ./data/hi/culturax \
    --repo-type dataset
```

The `run_pipeline.py` script automates this:
```bash
python run_pipeline.py --language hi --stage download
python run_pipeline.py --language tr --stage download
```

---

### Stage 2: Quality Filtering

Quality filtering uses DataTrove with custom language-specific filters.

#### Filter Configuration (from Fineweb-2)

**Hindi Filter Thresholds** (`src/filters/hi_quality.py`):
```python
HINDI_FILTER_CONFIG = {
    "min_script_ratio": 0.5,         # Devanagari script ratio
    "lang_score_threshold": 0.692,
    "dup_line_frac": 0.206,
    "new_line_ratio": 0.316,
    "min_avg_word_length": 2,
    "max_avg_word_length": 21,
    "line_punct_thr": 0.091,
    "non_alpha_words_ratio": 0.837,
    "top_5_gram_frac": 0.135,
    "top_10_gram_frac": 0.090,
}
```

**Turkish Filter Thresholds** (`src/filters/tr_quality.py`):
```python
TURKISH_FILTER_CONFIG = {
    "min_script_ratio": 0.65,        # Turkish Latin script ratio
    "lang_score_threshold": 0.875,
    "dup_line_frac": 0.272,
    "new_line_ratio": 0.222,
    "min_avg_word_length": 3,
    "max_avg_word_length": 21,
    "line_punct_thr": 0.091,
    "non_alpha_words_ratio": 0.773,
    "top_5_gram_frac": 0.154,
    "top_10_gram_frac": 0.103,
}
```

#### Output Schema Normalization

All filtered output uses a unified schema:
```python
OUTPUT_SCHEMA = pa.schema([
    ("text", pa.string()),
    ("id", pa.string()),
    ("metadata", pa.struct([
        ("source", pa.string()),
    ])),
])
```

#### Running Filtering

```bash
# Main parquet datasets (culturax, fineweb2, sangraha, vngrs)
python run_pipeline.py --language hi --stage filter
python run_pipeline.py --language tr --stage filter

# C4 JSON files (requires separate script due to JSON format)
python filter_c4.py --language hi
python filter_c4.py --language tr

# HPLT2 (requires separate handling due to nested structure)
python fix_hindi_hplt2_filter.py
python fix_hplt2_filter.py  # Turkish - 5 subfolders
```

---

### Stage 3: MinHash Deduplication

MinHash deduplication removes near-duplicate documents within each source.

#### MinHash Configuration
```python
MINHASH_CONFIG = {
    "n_grams": 5,
    "num_buckets": 14,
    "hashes_per_bucket": 8,
    "similarity_threshold": 0.8,
}
```

#### MinHash Stages (via DataTrove)
1. **Stage 1 - Signatures**: Generate MinHash signatures for each document
2. **Stage 2 - Buckets**: Group documents by LSH buckets to find candidates
3. **Stage 3 - Cluster**: Cluster similar documents together
4. **Stage 4 - Filter**: Keep one representative per cluster, write deduped output

#### Running MinHash
```bash
python run_pipeline.py --language hi --stage minhash
python run_pipeline.py --language tr --stage minhash
```

**Runtime**: Hindi ~17 hours, Turkish ~30 hours (on 56-core machine)

---

### Stage 4: Consensus Subset Construction

The consensus subset identifies documents that appear in 2+ sources using exact text hash matching.

#### Algorithm (Two-Pass, Memory-Efficient)

**Pass 1**: Build hash-to-sources index
```python
# For each document, compute MD5 hash of normalized text
# Store only: hash -> set of sources (not full text)
def compute_text_hash(text: str) -> str:
    normalized = ' '.join(text.lower().split())
    return hashlib.md5(normalized.encode('utf-8')).hexdigest()

# Pass 1: hash_to_sources[hash].add(source)
```

**Pass 2**: Extract documents with multi-source hashes
```python
# Re-read data, collect documents where hash appears in 2+ sources
# Store full document with sources list
```

#### Consensus Output Schema
```python
schema = pa.schema([
    ('text', pa.string()),
    ('id', pa.string()),
    ('sources', pa.list_(pa.string())),      # e.g., ["c4", "culturax"]
    ('all_ids', pa.list_(pa.string())),      # e.g., ["c4:url1", "culturax:url2"]
    ('metadata', pa.struct([
        ('source', pa.string()),              # "consensus"
    ])),
])
```

#### Running Consensus Builder
```bash
python build_consensus_v2.py --language hi
python build_consensus_v2.py --language tr
```

**Runtime**: Hindi ~2 hours, Turkish ~7 hours

---

### Stage 5: Upload to HuggingFace

#### Create Repositories
```bash
huggingface-cli repo create HinMix --organization AdaMLLab --type dataset
huggingface-cli repo create TurMix --organization AdaMLLab --type dataset
```

#### Staging Directory Structure
```
hf_staging/HinMix/
├── README.md
├── minhash_deduped/
│   ├── c4/*.parquet
│   ├── culturax/*.parquet
│   ├── fineweb2/*.parquet
│   ├── hplt2/*.parquet
│   ├── sangraha_unverified/*.parquet
│   └── sangraha_verified/*.parquet
├── quality_filtered/
│   └── (same structure as minhash_deduped)
└── consensus/
    └── consensus.parquet
```

#### Upload Command
```bash
# Use upload-large-folder for large datasets
hf upload-large-folder AdaMLLab/HinMix hf_staging/HinMix --repo-type dataset --num-workers 8
hf upload-large-folder AdaMLLab/TurMix hf_staging/TurMix --repo-type dataset --num-workers 8
```

---

## Final Statistics

### Hindi (HinMix)

| Stage | Documents | Size | Notes |
|-------|-----------|------|-------|
| **Quality Filtered** | ~99M | 231GB | All 6 sources combined |
| **MinHash Deduped** | ~60M | 136GB | 40% reduction |
| **Consensus** | 1.92M | 3.7GB | Docs in 2+ sources |

**Consensus Source Participation**:
- fineweb2: 1,602,172
- hplt2: 1,194,132
- sangraha_unverified: 600,851
- culturax: 277,990
- sangraha_verified: 153,060
- c4: 71,462

### Turkish (TurMix)

| Stage | Documents | Size | Notes |
|-------|-----------|------|-------|
| **Quality Filtered** | ~49M | 658GB | All 5 sources combined |
| **MinHash Deduped** | ~27M | 359GB | 46% reduction |
| **Consensus** | 7.84M | 13GB | Docs in 2+ sources |

**Consensus Source Participation**:
- fineweb2: 7,217,270
- hplt2: 7,075,189
- culturax: 686,152
- c4: 419,307
- vngrs: 402,760

---

## Key Scripts

### run_pipeline.py (Main Entry Point)
```python
#!/usr/bin/env python3
"""
Usage:
    python run_pipeline.py --language hi --stage download
    python run_pipeline.py --language hi --stage filter
    python run_pipeline.py --language hi --stage minhash
    python run_pipeline.py --language tr --stage all
"""
```

### build_consensus_v2.py (Consensus Builder)
```python
#!/usr/bin/env python3
"""
Memory-efficient two-pass consensus builder.
Usage:
    python build_consensus_v2.py --language hi
    python build_consensus_v2.py --language tr
"""
```

### filter_c4.py (C4 JSON Filtering)
```python
#!/usr/bin/env python3
"""
Filters C4 JSON files (not parquet) with schema normalization.
Usage:
    python filter_c4.py --language hi
    python filter_c4.py --language tr
"""
```

---

## Troubleshooting

### Common Issues

1. **Storage Full During MinHash**
   - MinHash signatures can grow to 300+ GB
   - Ensure at least 500GB free space before starting
   - If interrupted, clean `minhash_signatures/`, `minhash_buckets/`, `minhash_clusters/` and restart

2. **Memory Issues During Consensus**
   - Use `build_consensus_v2.py` (two-pass, memory-efficient)
   - Original `build_consensus.py` requires 100+ GB RAM

3. **C4 Schema Mismatch**
   - C4 is JSON, not parquet
   - Use `filter_c4.py` with `JsonlReader` and schema adapter

4. **HPLT2 Nested Folders**
   - Turkish HPLT2 has 5 subfolders (`tur_Latn_1` to `tur_Latn_5`)
   - Use `fix_hplt2_filter.py` which handles all subfolders

### Recovery Commands
```bash
# Check running processes
ps aux | grep "run_pipeline\|build_consensus" | grep -v grep

# Check disk usage
df -h /home/alrashsm/Documents/Github/arabic-pretraining-mix-other-languages/data/

# Check MinHash progress
ls data/hi/logs/minhash_sig/completions/ | wc -l  # Should reach 54
tail -20 data/hi/minhash.log
```

---

## License

This pipeline and resulting datasets are released under CC-BY-4.0.
Individual source datasets have their own licenses - refer to original sources.

---

## Citation

```bibtex
@dataset{hinmix_turmix_2024,
  title={HinMix and TurMix: Hindi and Turkish Pretraining Data Mixes},
  author={AdaMLLab},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/AdaMLLab}
}
```