Large-Scale Preference Dataset by Kaggle | Other

About this Dataset

Large-Scale Preference Dataset

Training Powerful Reward & Critic Models with Aligned Language Models

By Huggingface Hub [source]

About this dataset

UltraFeedback is an unprecedentedly expansive, meticulously detailed, and multifarious preference dataset built exclusively to train powerful reward and critic models with aligned language models. With thousands of prompts lifted from countless distinct sources like UltraChat, ShareGPT, Evol-Instruet, TruthfulQA and more, UltraFeedback contains an overwhelming 256k samples – perfect for introducing to a wide array of AI-driven projects. Dive into the selection of correct answers and incorrect answers attached to this remarkable collection easily within the same data file! Get up close in exploring options presented in UltraFeedback – a groundbreaking new opportunity for data collectors!

More Datasets

For more datasets, click here.

Featured Notebooks

🚨 Your notebook can be here! 🚨!

How to use the dataset

The first step is to understand the content of the dataset, including source, models, correct answers and incorrect answers. Knowing which language models (LM) were used to generate completions can help you better interpret the data in this dataset.

Once you are familiar with the column titles and their meanings it’s time to begin exploring! To maximize your insight into this data set use a variety of visualization techniques such as scatter plots or bar charts to view sample distributions across different LMs or answer types. Analyzing trends between incorrect and correct answers through data manipulation techniques such as merging sets can also provide valuable insights into preferences across different prompts and sources.

Finally, you may want to try running LR or other machine learning models on this dataset in order to create simple models for predicting preferences when given inputs from real world scenarios related to specific tasks that require nuanced understanding of instructions provided by one’s peers or superiors.

The possibilities for further exploration of this dataset are endless - now let’s get started!

Research Ideas

Training sentence completion models on the dataset to generate responses with high accuracy and diversity.

Creating natural language understanding (NLU) tasks such as question-answering and sentiment analysis using the aligned dataset as training/testing sets.

Developing strongly supervised learning algorithms that are able to use techniques like reward optimization with potential translation applications in developing machine translation systems from scratch or upstream text-generation tasks like summarization, dialog generation, etc

Acknowledgements

If you use this dataset in your research, please credit the original authors.
Data Source

License

License: CC0 1.0 Universal (CC0 1.0) - Public Domain Dedication
No Copyright - You can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission. See Other Information.

Columns

File: train.csv

Column name	Description
source	The source of the data. (String)
instruction	The instruction given to the language models. (String)
models	The language models used to generate the completions. (String)
correct_answers	The correct answers to the instruction. (String)
incorrect_answers	The incorrect answers to the instruction. (String)

Acknowledgements

If you use this dataset in your research, please credit the original authors.
If you use this dataset in your research, please credit Huggingface Hub.

Tables

Train

@kaggle.thedevastator_large_scale_preference_dataset.train

344.65 MB
63967 rows
6 columns


CREATE TABLE train (
  "source" VARCHAR,
  "instruction" VARCHAR,
  "models" VARCHAR,
  "completions" VARCHAR,
  "correct_answers" VARCHAR,
  "incorrect_answers" VARCHAR
);