toiar/khasi-instruction-response-v2

Khasi Instruction Response v2

The Khasi Instruction Response v2 dataset is a high-quality, curated collection of 77,810 instruction-response pairs designed to fine-tune Large Language Models (LLMs) for the Khasi language. This is an improved, expanded version of my previous v1 release, offering significantly higher data integrity and broader linguistic coverage.

It combines extensive cultural, literary, and translation-based Khasi data with high-reasoning capabilities from state-of-the-art models sourced from English. By integrating these English-source reasoning chains, the dataset enables the model to perform complex logical problem-solving while maintaining fluency in Khasi.

Dataset Statistics

• Total Conversations: 77,810
• Khasi-Specific Data: 72,810
• External Reasoning/General Data: 5,000

Uses

Direct Use

This dataset is intended for instruction-tuning LLMs (such as Gemma, Llama, or Mistral) to improve their proficiency in Khasi, cross-lingual translation, and logical reasoning.

Out-of-Scope Use

• Not intended for generating formal medical or legal advice.
• Not intended for use in high-stakes automated decision-making systems without human oversight.

Dataset Structure

The dataset is provided in both .jsonl and .parquet formats within the data/ directory. Each entry follows a standard ChatML schema:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "...", "reasoning": "..."}
  ]
}

Dataset Creation

Bias, Risks, and Limitations

• Linguistic Bias: As the dataset relies on collected literature and web-scraped data, it may reflect the dialectal preferences or formal writing styles dominant in those sources.
• Translation Bias: A significant portion of the Khasi data consists of English-to-Khasi translation pairs.

Citation

If you use this dataset in your research or project, please cite it as:

@dataset{khasi_instruction_response_v2,
  author = {Toiarbor Mawlieh},
  title = {Khasi Instruction Response v2},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{[https://huggingface.co/datasets/toiar/khasi-instruction-response-v2](https://huggingface.co/datasets/toiar/khasi-instruction-response-v2)}}
}