aghatage/marathi-alpaca-llama-finetune

Marathi Alpaca Dataset for llama-finetune

This dataset contains 48,897 high-quality Marathi instruction-following examples, converted to the llama-finetune format.

Dataset Description

• —Language: Marathi (मराठी)
• —Size: 48,897 examples
• —Format: JSONL with messages structure
• —Source: Converted from smallstepai/marathi-instruction-tuning-alpaca
• —Use Case: Supervised fine-tuning (SFT) for language models

Format

Each line in the JSONL file contains:

{
  "messages": [
    {
      "role": "user",
      "content": "निरोगी राहण्यासाठी तीन टिपा द्या."
    },
    {
      "role": "assistant",
      "content": "1. संतुलित आणि पौष्टिक आहार घ्या..."
    }
  ]
}

Usage

Download

from datasets import load_dataset

dataset = load_dataset("aghatage/marathi-alpaca-llama-finetune")

With llama-finetune

# Download the JSONL file
wget https://huggingface.co/datasets/aghatage/marathi-alpaca-llama-finetune/resolve/main/marathi_alpaca_finetune.jsonl

# Use with llama-finetune
./llama-finetune \
    --model-base lfm2-350m \
    --train-data marathi_alpaca_finetune.jsonl \
    --output-dir ./output

Data Quality

The dataset is translated from the high-quality Alpaca-Cleaned dataset, ensuring:

• —✅ Diverse instruction-following tasks
• —✅ Clean and well-formatted data
• —✅ Suitable for supervised fine-tuning
• —✅ Covers general knowledge and reasoning tasks

Original Dataset

This dataset is a format conversion of smallstepai/marathi-instruction-tuning-alpaca, which is a Marathi translation of the Alpaca dataset.

Citation

If you use this dataset, please cite the original Alpaca and Marathi translation work:

@misc{marathi-alpaca-2024,
  author = {SmallStepAI},
  title = {Marathi Instruction Tuning Alpaca Dataset},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/smallstepai/marathi-instruction-tuning-alpaca}
}

License

Apache 2.0 - Same as the original dataset

Conversion

This dataset was converted using the conversion script available at the repository.