Models
GooseAI offers out-of-the-box open source model endpoints that are ready to use. Models can be vastly different, in the type and size of the information set they were trained using, and while we’ve written basic summaries on each, we strongly recommend you do testing both in the GooseAI Playground as well as testing in your product to find the best model and parameter fit for your use case.
The models that GooseAI currently supports are:
GPT NeoX 20B
NeoX 20B is the latest model produced by EleutherAI and the biggest open sourced Language Model.
API Model Name:
gpt-neo-20b
Dataset Used: The Pile
Fairseq Series
Fairseq models are trained by Meta research labs as a reproduction of GPT-3.
They are our recommended models for their respective sizes and perform better than their GPT-Neo counterparts in most cases.
Fairseq models are especially good for Classification, but they can be used for all kinds of tasks.
API Model Names:
fairseq-125m
fairseq-1-3b
fairseq-2-7b
fairseq-6b-7b
fairseq-13b
Dataset Used: Fairseq is trained on filtered Common Crawl, Books and Wikipedia entries.
GPT-J 6B
GPT-J is a great open source alternative for the mid-size range of GPT-3 models. Trained by EleutherAI, it's used extensively for text generation and completion use cases.
API Model Name:
gpt-j-6b
Dataset Used: The Pile
GPT-Neo Series
GPT-Neo models are trained by EleutherAI as a reproduction of GPT-3.
API Model Names:
gpt-neo-125m
gpt-neo-1-3b
gpt-neo-2-7b
Dataset Used: The Pile