botbots

The botbots dataset is an open-source collection of conversations between two ChatGPT (gpt-3.5-turbo) instances, published by radi-cho under the MIT License. The CLI commands and dialogue prompts were written by GPT-4, and the texts were generated with datasetGPT using the OpenAI API as the backend. It contains 2361 conversations with 29796 utterances.

What It Is

botbots is a dataset for researchers and developers working with conversational language models. Each sample is a JSON record with the system prompts for both agents, generation settings such as temperature and length, and the list of utterances. The README suggests two uses: studying the inventive, adaptability and logical abilities of LLMs (especially gpt-3.5-turbo), and training smaller conversational models in an Alpaca-like fashion.

Dataset Contents

The data is split into three groups:

• Task-oriented dialogues: 1461 conversations with 20784 utterances across 16 domains, including hotels, restaurants, medical appointments, event tickets, flight bookings, and tech support. Situation-specific values are labelled, for example [hotel|The Cursed Castle], which supports entity recognition and state tracking.
• Reasoning: 800 conversations with 8001 utterances in 10 settings, such as spatial, logical, inductive, deductive, causal, counterfactual, and strategic reasoning.
• Brainstorming: 100 conversations with 1011 utterances on 25 scientific problems, including climate change, quantum computing, and neuroscience.

How It Was Generated

A datasetGPT command for hotel dialogues was designed manually. GPT-4 was then prompted to produce similar commands for other domains and to extend them to reasoning and brainstorming conversations. These commands are kept in the .gpt4.txt files, of which brainstorming.gpt4.txt is described as the most insightful. Some prompts were reused across temperature settings, so questions or conversation fragments may repeat; the authors consider these duplicates useful for evaluating the diversity of LLM outputs.