license: cc-by-sa-3.0
task_categories:
- question-answering
language:
- en
Transformed Databricks-Dolly-15k Dataset
Summary
The Transformed Databricks-Dolly-15k dataset is a modification of the original open-source dataset created by Databricks employees, designed to facilitate instruction-following abilities in large language models (LLMs). This version has been specifically adapted to include responses in a JSON format, enhancing its utility for tasks requiring structured output.
Modifications
The primary transformation applied to the original dataset involves converting the response field of each record into a JSON object with a key "answer", making it directly usable for fine-tuning LLMs with a preference for JSON-formatted outputs.
Supported Tasks
- Training LLMs with structured JSON outputs
- Synthetic Data Generation with JSON responses
- Data Augmentation for structured data applications
Languages
English
Version
Transformed Dataset: 1.0 Based on Original Dataset Version: 1.0
Owner
Databricks, Inc.
Dataset Overview
The dataset comprises over 15,000 records, with each entry now featuring a response encapsulated in a JSON object. These modifications are intended to support the development of LLMs capable of generating structured outputs, facilitating a wide range of applications from synthetic data generation to data augmentation.
Intended Uses
This transformed dataset is suited for the same broad range of applications as the original, with the added benefit of structured responses for tasks that require or benefit from JSON-formatted output.
Sources
- Human-generated data: Created by Databricks employees.
- Wikipedia: Used for certain categories requiring reference texts.
Annotator Guidelines
The original dataset's annotation guidelines remain relevant, with the additional step of structuring responses in JSON format.
Personal or Sensitive Data
No personal or sensitive data is included in this dataset.
Language
American English
Known Limitations
The limitations of the original dataset, including potential biases and factual errors inherent in Wikipedia, are applicable to this transformed version.
Citation
Please cite the original dataset as follows, with an additional note on the transformation if necessary:
@online{DatabricksBlog2023DollyV2,
author = {Mike Conover and Matt Hayes and Ankit Mathur and Jianwei Xie and Jun Wan and Sam Shah and Ali Ghodsi and Patrick Wendell and Matei Zaharia and Reynold Xin},
title = {Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM},
year = {2023},
url = {https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm},
urldate = {2023-06-30}
}
License/Attribution This transformed dataset is based on work developed at Databricks (https://www.databricks.com) and its use is subject to the Creative Commons Attribution-ShareAlike 3.0 Unported License, consistent with the original dataset's licensing terms.