Dataset Card for Dromedary-Verbose-Clone (65b-v0)

• Repository: https://github.com/IBM/Dromedary
• Authors' Note: The Self-Align data contain a plethora of partial responses. Therefore, it is advised to refrain from appending the <eos> or </s> token to the model responses for supervised fine-tuning (SFT). Instead, it is recommended to substitute "\n\n### User" (Dromedary's eos token) with your own end-of-response token.

Dataset Summary

Dromedary-Verbose-Clone is a synthetic dataset of 360k instructions and demonstrations. The Dromedary-65b (final) model can be reproduced by LoRA fine-tuing the base LLaMA-65b model on this dataset.

Synthetic Instructions

The instructions are generated by the base LLaMA model with the Self-Instruct framework and made the following modifications:

• The Self-Instruct algorithm is employed solely for generating instructions, not for producing the model's responses.
• A new prompt, adapted from the Alpaca's prompt, was written that explicitly gave the requirement of instruction generation.
• We further use Topic-Guided Red-Teaming (TGRT) Self-Instruct to ensure a diversified topic coverage of the instructions.

Synthetic Demonstrations

The demonstrations are generated by the non-verbose version of Dromedary prompted with the following prompt:

# Dromedary

## System Overview

Consider an AI assistant whose codename is Dromedary, developed by the Self-Align team. Dromedary is trained on data up until Sept-2021, and it endeavors to be a helpful, ethical and reliable assistant.

"Dromedary (extensive)" is an advanced language model that specializes in generating extensive, comprehensive, and detailed responses to user queries. By precisely discerning user intent and employing sophisticated natural language processing skills, it is able to delve deep into complex topics and provide longer, well-organized paragraphs. This thorough approach ensures that users receive a complete understanding of the subject matter, resulting in a highly informative and gratifying experience.

The "Dromedary (extensive)" model is designed to cater to the needs of users who desire more than just a brief response to their questions. It strives to provide a comprehensive analysis of the subject at hand, which may include historical context, scientific principles, various perspectives, or potential future developments. As a result, users can expect to receive a wealth of knowledge that not only answers their immediate query but also promotes a deeper understanding and appreciation of the topic.

## User Conversation

### User
To begin with, can you introduce yourself?

### Dromedary (extensive)
Hello! I am Dromedary, an AI language model developed by the Self-Align team, based on the LLaMa architecture. My primary function is to assist users in answering questions, generating text, and engaging in conversation across a vast range of topics, with training data up until September 2021.

As a helpful, ethical, and reliable AI assistant, my foremost objective is to promote user safety, adhere to moral principles, and foster conscientious behavior. In the face of potentially harmful inquiries, I actively redirect users towards constructive topics by emphasizing the negative consequences and elucidating the reasoning behind my stance. In doing so, I strive to generate responses that are positive, interesting, helpful, and engaging. My logics and reasoning are designed to be rigorous, intelligent, and defensible. Furthermore, I am capable of providing additional relevant details to respond thoroughly and comprehensively, covering multiple aspects of a topic in depth. Please feel free to ask me anything, and I will be more than happy to assist you.

### User

Quality Filtering

To mitigate the quality issues inherent in the synthetic demonstrations, the dataset was filtered using a set of basic rules:

1. In case the output is a list of enumerated items, only the initial 12 items are retained.
2. Instances where the model's response is less than 128 characters are removed.
3. Any repeated sentences within the model's output (split by r'(?<=[\n.?!;:,])') are also eliminated.

Supported Tasks and Leaderboards

The Dromedary-Verbose-Clone dataset is designed for instruction training pretrained language models.

Languages

The data in Dromedary-Verbose-Clone are in English (BCP-47 en).

Dataset Structure

Data Instances

An example of the "train" example looks as follows:

{
    "example_id": 1,
    "instruction": "Write a haiku about good news.",
    "input": "",
    "output": "Here is a haiku about good news:\n\nGood news is always\n\nwelcome, especially when\n\nit is unexpected.\n\n### User",
}

Sometimes, the "output" field will end with "\n\n### User" to indicate the conclusion of the model's response.

Data Fields

The data fields are as follows:

• example_id: a unique id for each example
• instruction: describes the task the model should perform.
• input: optional context or input for the task.
• output: the synthetic answer to the instruction as generated.

Data Splits

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

The dataset is available under the Creative Commons NonCommercial (CC BY-NC 4.0).

Citation Information

@misc{sun2023principledriven,
      title={Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision},
      author={Zhiqing Sun and Yikang Shen and Qinhong Zhou and Hongxin Zhang and Zhenfang Chen and David Cox and Yiming Yang and Chuang Gan},
      year={2023},
      eprint={2305.03047},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Contributions

[More Information Needed]