YoseAli/amharic-llm-training-data

Amharic LLM Training Dataset

Complete production-ready Amharic dataset for large language model training and deployment.

🚀 Quick Start for Deployment

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("YoseAli/amharic-llm-training-data")

# Access splits
train_data = dataset["train"]  # 761,501 samples
test_data = dataset["test"]    # 84,612 samples

print(f"Training samples: {len(train_data):,}")
print(f"Test samples: {len(test_data):,}")

📊 Dataset Information

• Total samples: 846,113
• Training samples: 761,501
• Test samples: 84,612
• Language: Amharic (am)
• Format: Instruction-response pairs
• Quality: Curated and validated
• Sources: Multi-source compilation (Walia-LLM, AYA, M2Lingual, AmQA, Masakhane)

🎯 Production Deployment

This dataset is optimized for:

• ✅ LLM Fine-tuning: Ready for transformer model training
• ✅ Production Deployment: Validated and production-ready
• ✅ Scalable Training: Supports distributed training
• ✅ Quality Assured: Curated from multiple high-quality sources

💻 Usage Examples

Basic Loading

from datasets import load_dataset

# Load dataset
dataset = load_dataset("YoseAli/amharic-llm-training-data")
training_data = dataset["train"]

# Convert to list for custom processing
train_list = training_data.to_list()

Training Format

# Format for instruction tuning
def format_for_training(example):
    if 'instruction' in example and 'output' in example:
        text = f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"
    elif 'text' in example:
        text = example['text']
    else:
        text = str(example)
    
    return {"text": text}

# Apply formatting
formatted_dataset = dataset.map(format_for_training)

Streaming for Large Scale

# For very large scale training
dataset = load_dataset("YoseAli/amharic-llm-training-data", streaming=True)
train_stream = dataset["train"]

# Process in batches
for batch in train_stream.iter(batch_size=1000):
    # Your training code here
    pass

🏗️ Model Training Pipeline

1. Data Preparation

from datasets import load_dataset
from transformers import AutoTokenizer

# Load dataset
dataset = load_dataset("YoseAli/amharic-llm-training-data")
tokenizer = AutoTokenizer.from_pretrained("your-base-model")

# Tokenize
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, padding=True)

tokenized_dataset = dataset.map(tokenize_function, batched=True)

2. Training Setup

from transformers import AutoModelForCausalLM, TrainingArguments, Trainer

# Load model
model = AutoModelForCausalLM.from_pretrained("your-base-model")

# Training arguments
training_args = TrainingArguments(
    output_dir="./amharic-model",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-5,
    save_steps=500,
    logging_steps=100,
)

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

# Start training
trainer.train()

📋 Data Format

Each sample contains:

{
  "instruction": "Task or question in Amharic",
  "input": "Additional context (optional)",
  "output": "Expected response in Amharic",
  "source_dataset": "Original source (when available)"
}

🌍 Dataset Sources

This dataset combines multiple high-quality Amharic language resources:

1. Walia-LLM: Amharic instruction-following dataset
2. AYA Amharic: Cohere's multilingual dataset (Amharic subset)
3. M2Lingual: Cross-lingual dialogue dataset
4. AmQA: Amharic question-answering from Wikipedia
5. Masakhane NLU: African languages NLU benchmark

🚀 Deployment Ready

This dataset is ready for:

• Production LLM training
• Commercial applications
• Research and development
• Community model building
• Educational purposes

📈 Performance Expectations

Models trained on this dataset typically achieve:

• Strong Amharic language understanding
• Good instruction-following capabilities
• Diverse response generation
• Cultural and contextual awareness

📄 Citation

If you use this dataset in your work, please cite:

@dataset{amharic_llm_training_data,
  title={Amharic LLM Training Dataset},
  author={YoseAli},
  year={2024},
  url={https://huggingface.co/datasets/YoseAli/amharic-llm-training-data},
  note={Complete Amharic LLM training dataset from multiple curated sources}
}

📜 License

MIT License - Free for research and commercial use.

🤝 Acknowledgments

• EthioNLP community for Walia-LLM dataset
• Cohere for AYA multilingual dataset
• Masakhane community for African NLP resources
• All contributors to the original source datasets

📞 Contact

For questions, issues, or collaboration:

• Repository: GitHub
• Hugging Face: @YoseAli

---

Ready for production deployment! 🚀