Orca AgentInstruct 1M is a synthetic instruction dataset generated based on the AgentInstruct framework, released by Microsoft in 2024. The related paper is "AgentInstruct: Toward Generative Teaching with Agentic Flows", and it aims to provide high-quality instruction fine-tuning data for large language models.

The dataset contains approximately 1 million synthetic instruction pairs, with all prompts and responses generated using only publicly available raw text content as seeds via the AgentInstruct framework. The Orca-3-Mistral model, obtained by post-training Mistral-7b on the full AgentInstruct dataset (approximately 25 million instruction pairs), achieved significant improvements over the original model across multiple benchmarks such as AGIEval, MMLU, and GSM8K. The data covers various capability dimensions including text editing, creative writing, programming, and reading comprehension, and can be used for instruction fine-tuning research on any base large model.

Dataset Composition

• creative_content: Creative content generation.
• text_modification: Text modification and polishing.
• struct2text_flow: Structured data to text conversion.
• rc: Reading comprehension.
• rag: Retrieval-augmented generation.
• text_extraction: Text information extraction.
• mcq: Multiple-choice questions.
• follow_up: Follow-up question generation and answering.
• analytical_reasoning: Analytical reasoning.
• fermi: Fermi estimation problems.
• fs_cot_flow: Few-shot chain-of-thought reasoning.
• code: Code generation and programming.
• brain_teaser: Brain teasers and logic puzzles.
• text_classification: Text classification.
• open_domain_qa: Open-domain question answering.

Citation

@misc{  
      title={ AgentInstruct: Toward Generative Teaching with Agentic Flows},  
      author={Arindam Mitra, Luciano Del Corro, Guoqing Zheng, Shweti Mahajan, Dany Rouhana, Andres Codas, Yadong Lu, Wei-ge Chen, Olga Vrousgos, Corby Rosset, Fillipe Silva, Hamed Khanpour, Yash Lara, Ahmed Awadallah}, 
      year={2024}, 
      eprint={ 2407.03502}, 
      archivePrefix={arXiv}, 
      primaryClass={cs.CL} 
}