Check out atlas/scripts/example.py for example code.
To run:
cd atlas
python3 scripts/example.py
This script demonstrates getting tokenization and token counts, few shot completions, and staged submitting. By default the code uses a mock version of GPT. To submit the queries for real with the existing scripts, append "submit" to the command; E.g. python3 scripts/example.py submit.
Percy's process.py script provides a wrapper around the official OpenAI API with two convenience functions, .complete() for language generation and .few_shot() for few_shot learning experiments with optional labels. Both of these call OpenAI's openai.Completion.create underneath.
Every request via openai.Completion.create will be cached in a cache<...>.jsonl file so that subsequent identical requests will be loaded from the cache instead of sent to the server again.
The code under atlas is a second implementation, with atlas.gpt.py being the analog of process.py. atlas includes code that tries to take a more functional approach through atlas/dataset.py as demoed in atlas/sequence_manipulation.py, which reeks of computation graphs. Other files of possible interest include atlas/api.py and atlas/gpt2.py.
For a more interactive way of querying the API that will still cache your queries, check out atlas/scripts/interactive.py.
There is a slack bot that you can query for API usage under the name /lassie.
Getting the logprobs (max 100) can result in large caches. If you're running into this and want both to get the logprobs as well as have fast loading when you don't need all the logprobs, atlas/scripts/reformat_cache_logprobs.py gets rid of all but the top n=1 predicted tokens after each token and saves the modified dictionary to <cache_file_name>_lps-1.jsonl.
API documentation for the openai.Completion.create function is here:
and copy+pasted here with a few additional options like frequency_penalty and presence_penalty for reducing repetition.
There are four main engines, in order of smallest to largest: ada, babbage, curie, davinci. By default, the engine in process.py is set to davinci (see DEFAULT_GENERATION_KWARGS).
Please note that this repository is unstable (in developer terms)! Send questions and comments to @friedeggs (Frieda).
[2021-01-11] Committed example code.
[2020-10-07] Added staging of requests so that when submitting requests with staged=True, those requests will not be run until the user calls gpt3.run_staged_queries() at which point they can see how many requests are queued in total, choose to run all requests (y), run with confirmation for each request (c), or quit without submitting. When processing staged requests, a progress bar will indicate how many requests are left.