Simulate people with GPT models and generate response statistics. Not sure how to do any of these things yet but we're working on it.
A simulated survey is generated from a prompt file in ./prompts/
directory.
Prompt file should consist by a series of prompts separated by ---
.
For example, a prompt file:
prompt1
---
prompt2
---
prompt3
will result in the following conversation:
user: prompt1
assistant: response1
user: prompt2
assistant: response2
user: prompt3
assistant: response3
Use a prompt file prompt_name.txt
to simulate a survey with num
participants via the command
python -m collect_responses -p prompt_name -n num
The default values for the arguments are prompt_name=baseline
and num=10
.
A survey will be saved in ./data/prompt_name/
as a series of .json
files numbered after each simulated participant.