gpt-3-experiments

A repo containing test prompts for OpenAI's GPT-3 API and the resulting AI-generated texts, which both illustrate the model's robustness, plus a Python script to quickly query texts from the API. All generated texts in this repo are completely unedited and uncurated unless explicitly stated otherwise.

Disclaimer: generated text content in this repository may be offensive. The READMEs of the corresponding examples will include an explicit content warning (CW) when this is the case.

Repo Layout

This repo contains folders for each prompt example in the /examples folder. The README for each prompt example contains the input and any content warnings as noted above.

All texts were generated from the best davinci model. Specifically, after feeded the prompt, 1 text was generated (at up-to-512 tokens per text) at temperature=0.0 (i.e. the model will always choose the most likely output and is therefore deterministic), and 10 texts for each temperature of 0.7, 1.0, and 1.2: the higher the temperature, the more "creative" the text.

Script Usage

If you have access to the OpenAI API, you can use the openai_api.py Python script to gather generated texts faster than using the web interface.

First, download/clone the repo, and in the config.yml, set the SECRET_KEY to the one provided to you by the OpenAI API. (do not share with anyone!)

The script (Python 3.6+) requires the installation of certain Python packages:

pip3 install httpx pyyaml fire tqdm

After that, you can run it from the command line. For example, if you wanted to generate text with the prompt "Once upon a time", you could do:

python3 openai_api.py "Once upon a time"

This will generate a file for each specified temperature in the config.yml.

If you want to read a longer text from a file (e.g. prompt.txt, the default behavior), you can put that prompt in that file and run:

python3 openai_api.py "prompt.txt"

By default, the output files are Markdown files, which allows them to include the prompt bolded and render better on GitHub. To save the output files as text instead (w/o the prompt), set:

python3 openai_api.py "prompt.txt" --markdown False

Notes

• The script uses synchronous requests by default with a 30 second sleep between requests: apparently the async approach which requested all generated texts simultaneously caused too much of a strain on OpenAI's infrastructure. Please uses that default for the time being.

Maintainer/Creator

Max Woolf (@minimaxir)

Max's open-source projects are supported by his Patreon and GitHub Sponsors. If you found this project helpful, any monetary contributions to the Patreon are appreciated and will be put to good creative use.

License

MIT

Disclaimer

This repo has no affiliation with OpenAI.