This is a test of converting the architecture of GPT-J 6B into a mixture-of-experts model. It is initialized with 4 experts (2 active) with otherwise the same configuration as GPT-J-6B, making this a 17B parameter model in total.
The model weights were initialized randomly - not loaded from the pretrained GPT-J - for testing purposes. This model is not useable for any downstream purposes unless you're trying to generate absolute schizo babble - in which case, this model is perfect for your use-case. You have been warned.
Be sure to pass trust_remote_code=True into AutoModelForCausalLM.from_pretrained if you still want to use this model for some god-forsaken reason.