Unnatural Instructions

This repository contains the Unnatural Instructions dataset. Unnatural Instructions is a dataset of instructions automatically generated by a Large Language model. See full details in the paper: "Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor"

🗃️ Content

The data folder contains two files: core_data.jsonl, containing the Unnatural Instructions core dataset of 68,478 instruction-input-output triplets, and full_data.jsonl, containing the full 240,670 Unnatural Instructions examples. The full data was constructed by expanding the core data with automatically generated instruction paraphrases.

📄 Format

Core data

Each line in core_data.jsonl is a JSON object with two fields - instruction, which is a natural language instruction describing a task, and instances, an array of JSON objects, each contains

• input: An input for the task described by the instruction
• instruction_with_input: The instruction concatenated with the input
• constraints: The task's output space constraints
• output: The output of executing instruction with the given input

Full data

core_data.jsonl has the same structure as core_data.jsonl, but with one additional field - reformulations. reformulations is an array of JSON objects, each corresponds to an automatically generated paraphrase for the given instruction. Each reformulation contains the fields:

• instruction: A paraphrase of the original instruction
• input: An input for the task described by the instruction
• instruction_with_input: The paraphrased instruction concatenated with the input
• output: The output of executing instruction with the given input

📘 Citation

If you make use of Unnatural Instructions, please cite the following paper:

@misc{honovich2022unnatural,
      title = {Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor},
      author = {Honovich, Or and Scialom, Thomas and Levy, Omer and Schick, Timo},
      url = {https://arxiv.org/abs/2212.09689},
      publisher = {arXiv},
      year={2022}
}