Databricks Dolly
15K Instruction Dataset
A high-quality instruction-following dataset of 15,000 examples written by Databricks employees, covering 7 major task categories, licensed under CC BY-SA 3.0, designed for fine-tuning large language models.
Dataset Highlights
High-quality human-written instruction data to assist in fine-tuning large language models
Human-Written Quality
All instruction-response pairs are manually written by Databricks employees, ensuring high quality and diversity of data, far exceeding AI-generated synthetic data.
Seven Task Categories
Covers seven core NLP task types: brainstorming, classification, closed-ended Q&A, text generation, information extraction, open-ended Q&A, and summarization.
No AI-Generated Content
The dataset does not contain any content generated by GPT or other large models, avoiding common data contamination issues in autoregressive training.
Commercially Friendly License
Released under the CC BY-SA 3.0 license, allowing commercial use and modification, making it an ideal training data source for building commercial LLMs.
Diverse Topics
Covers a wide range of topics including technology, history, geography, culture, and science, ensuring the fine-tuned model has comprehensive knowledge coverage.
Professional Author Team
Contributed by over 5,000 Databricks employees, all authors are technical professionals, ensuring the quality of instructions and the accuracy of responses.
Applicable Scenarios
From model training to quality benchmarks, meeting various LLM development needs
Instruction Fine-tuning
Directly used as LLM instruction fine-tuning data, quickly improving the model's instruction-following ability and dialogue quality
Task-Specific Training
Filtering data by 7 major categories for specialized model training on specific tasks such as summarization, classification, and Q&A
Quality Benchmark
As a quality reference benchmark for human-written data, evaluating and comparing the instruction-following quality of AI-generated data
Commercial LLM Development
Using the commercially friendly CC BY-SA 3.0 license to safely and compliantly build and deploy commercial large language models
Data Preview
The following are example records from the Dolly 15K dataset, showcasing the instruction-context-response structure
{
"instruction": "When was the first Star Wars movie released?",
"context": "",
"response": "The first Star Wars movie was released on May 25, 1977.",
"category": "closed_qa"
}
{
"instruction": "Give me a list of 5 ways to stay healthy.",
"context": "",
"response": "1. Exercise regularly\n2. Eat a balanced diet\n3. Get enough sleep\n4. Stay hydrated\n5. Manage stress effectively",
"category": "brainstorming"
}
{
"instruction": "Extract the country names from the following text.",
"context": "The United Nations has 193 member states, including France, Japan, Brazil, and Nigeria.",
"response": "France, Japan, Brazil, Nigeria",
"category": "information_extraction"
}
API Quick Call
Quickly obtain Dolly 15K data through the API, supporting category filtering
curl -X GET "https://api.acedata.cloud/datasets/dolly-15k" \ -H "Authorization: Bearer YOUR_API_TOKEN" \ -H "Content-Type: application/json"
3 Steps to Get Started Quickly
From obtaining the Token to loading data, you can start model training in just a few minutes
Obtain API Token
Register an account on the Ace Data Cloud platform and create an API Token to start calling the dataset interface.
Call Data Interface
Use the API to obtain data as needed, supporting filtering records by task category (e.g., brainstorming, closed_qa).
Start Model Training
Load the data into the training framework and use the instruction-response format for LLM instruction fine-tuning.
Start Using the Dolly 15K Dataset
High-quality manually crafted instruction data, commercially friendly license, available immediately. Whether you are fine-tuning an open-source model or building a commercial LLM, this dataset is an ideal starting point.