muraliwebworld/videxpulse-tamil-wiki-qa-200k

Tamil Language QA AI Model

An open-source project to build a Question-Answering system for Tamil Wikipedia content using a 3B-parameter language model fine-tuned with retrieval-augmented generation (RAG).

🎯 Quick Overview

This project scrapes Tamil Wikipedia, fine-tunes a lightweight language model, and deploys it as a RAG-based QA system that can answer questions about Tamil topics.

Key Features:

• 🔍 Automatic scraping of ~200K Tamil Wikipedia articles
• 🧠 3B-parameter model fine-tuned via LoRA and 4-bit quantization
• ⚡ FAISS-based retrieval for fast context lookup (<2s)
• 🎤 Interactive Streamlit web interface + FastAPI REST API
• 📦 Fully containerized with Docker
• 🔄 RAG architecture for context-aware answers

📊 Architecture

Wikipedia Tamil Data
        ↓
    [Scraper]
        ↓
  Cleaned Dataset
        ↓
   ┌─────────────────────────────────┐
   │ Generate Embeddings (FAISS)      │
   └─────────────────────────────────┘
        ↓
   [Fine-tune Model]
   (LoRA + 4-bit Quantization)
        ↓
   ┌─────────────────────────────────┐
   │ RAG System (Retrieval + Gen)     │
   │  - Retrieve relevant chunks      │
   │  - Generate Tamil answers        │
   └─────────────────────────────────┘
        ↓
   [Streamlit App / FastAPI]

� Dataset on HuggingFace Hub

We've published a clean, structured dataset on HuggingFace Hub:

Dataset: muraliwebworld/videxpulse-tamil-wiki-qa-200k

Dataset Statistics

• Format: Parquet (optimized for fast loading)
• Total Samples: 402,904 QA pairs
  • Train: 322,537 pairs (80%)
  • Validation: 40,423 pairs (10%)
  • Test: 39,944 pairs (10%)
• Language: Tamil (தமிழ்)
• Columns:
  • instruction: Tamil question/prompt
  • input: Empty (instruction-only format)
  • output: Answer/content from Wikipedia
  • source: Referenced Wikipedia links (comma-separated)
  • chunk_url: Original Wikipedia article URL
  • topic: Hierarchical topic path (Article - Section - Subsection)
  • language: Language code ("ta" for Tamil)

Load Dataset in Python

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("muraliwebworld/videxpulse-tamil-wiki-qa-200k")

# Load specific split
train_data = load_dataset("muraliwebworld/videxpulse-tamil-wiki-qa-200k", split="train")
val_data = load_dataset("muraliwebworld/videxpulse-tamil-wiki-qa-200k", split="validation")
test_data = load_dataset("muraliwebworld/videxpulse-tamil-wiki-qa-200k", split="test")

# Stream without downloading
dataset = load_dataset(
    "muraliwebworld/videxpulse-tamil-wiki-qa-200k",
    split="train",
    streaming=True
)

# Access individual samples
sample = dataset["train"][0]
print(f"Instruction: {sample['instruction']}")
print(f"Answer: {sample['output'][:100]}...")
print(f"Topic: {sample['topic']}")
print(f"Source: {sample['source']}")

Use in Fine-tuning (Transformers/SentenceTransformers)

from datasets import load_dataset
from sentence_transformers import InputExample, models, losses, SentenceTransformer
from torch.utils.data import DataLoader

# Load dataset
dataset = load_dataset("muraliwebworld/videxpulse-tamil-wiki-qa-200k", split="train")

# Convert to SentenceTransformer format
train_examples = [
    InputExample(
        texts=[item["instruction"], item["output"]],
        label=1.0  # High similarity for Q-A pairs
    )
    for item in dataset
]

# Use in training
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=16)

model = SentenceTransformer("sentence-transformers/distiluse-base-multilingual-cased-v2")
train_loss = losses.CosineSimilarityLoss(model)

# Fine-tune...

Dataset Features

✅ High Quality: Cleaned wiki markup, proper instruction format
✅ Hierarchical Topics: Preserves Wikipedia section hierarchy
✅ Source Attribution: Links to original Wikipedia articles
✅ Parquet Format: 3x faster loading than JSONL
✅ Instruction-Tuning Ready: Perfect for fine-tuning language models
✅ Multilingual: Pure Tamil content from Wikipedia
source venv/bin/activate # On Windows: venv\Scripts\activate

📝 License

• Code: MIT
• Data: Wikipedia Tamil (CC-BY-SA 3.0)
• Model: Depends on base model (see model cards)

🙏 Acknowledgments

• Tamil Wikipedia community
• HuggingFace ecosystem (transformers, datasets, spaces)
• FAISS library (Meta AI)
• Sentence-transformers library

❓ FAQ

Q: How much GPU memory do I need for training?
A: 8-16GB VRAM recommended. With LoRA + 4-bit quantization, can work with 6GB. Use Google Colab Pro or cloud GPUs.

Q: Can I use a different base model?
A: Yes! The pipeline is agnostic. Try Phi-2, Llama-2-7B (quantized), or Tamil-specific models if available.

Q: How do I deploy this in production?
A: See DEPLOYMENT.md for Docker, Kubernetes, and cloud options.

Q: Can I fine-tune on my own data instead of Wikipedia?
A: Yes! Replace the scraper with your data source. Format as JSON QA pairs and follow Phase 1 preprocessing.

Q: How do I improve answer quality?
A: See ARCHITECTURE.md for strategies: retrieval re-ranking, ensemble models, human feedback.

📞 Support

• Questions? Check the relevant guide (SETUP.md, TRAINING.md, DEPLOYMENT.md)
• Issues? See troubleshooting sections in each guide
• Ideas? Open a GitHub issue or contribute!

---

Status: 🚀 Ready for implementation
Last Updated: 2026-08-14
Maintained by: Community