Evol-character Character Setting and Dialogue Dataset

This dataset is generated by GPT3.5 and GPT4. To ensure the reasonable use of the data, only part of the data is currently open. The public data consists of three files, each of which contains the settings and dialogues of 200 characters.

Data Structure

• evol-character-gpt3.5.json
• evol-character-male-gpt3.5.json
• evol-character-gpt4.json

The details are as follows:

1. evol-character-gpt3.5.json: This dataset includes 200 different characters. The data of each character is divided into two parts: instruction and dialog. The instruction part describes the character's personality, experience and other characteristics, while the dialog part contains 10 groups of dialogues (but some characters may have less than 10 groups due to post-processing).
2. evol-character-male-gpt3.5.json: Also contains 200 characters, and its data structure is the same as evol-character-gpt3.5.json.
3. evol-character-gpt4.json: It also contains 200 characters, and the data is more detailed and sophisticated than gpt3.5 version. The data of each character is divided into two parts: setting and iqa. The setting part describes the character's personality, experience and other characteristics in detail, while the iqa part includes the personality settings of the characters who talk to the character, as well as multiple rounds of conversations between them. The data of each character covers three related characters and their conversations with the character.

Advantages

• Refined character setting data：This dataset makes up for the common problem of insufficient role settings in existing open source Role-playing Instruction data. It provides detailed information from multiple dimensions such as role identity, language style, and background story. Especially in the GPT-4 version, the dataset also adds the setting of the interlocutor's identity, making the data more complete and rich.
• Diverse character traits：This dataset covers the widest possible range of ACG character personalities, ensuring low duplication and high richness.
• Vivid language and action descriptions: This dataset not only contains the dialogues between characters, but also adds descriptions of the characters' actions, making the dialogues more vivid and realistic, which will provide users with a richer role-playing experience.
• Generic role-playing data generation framework：This dataset provides a general role-playing data generation framework to fully unleash the role-playing capabilities of the OpenAI API. The data generated by this framework will be used for fine-tuning and RAG. Currently, the framework code is being tested and optimized and is expected to be made public in the near future.