Extending Llama-3 8B with 1-million context by Gradient

Extending Llama-3 8B with 1-million context by Gradient
Photo by Willian Justen de Vasconcellos / Unsplash

Wow, things are heating up on this latest Meta model. I was looking to turn things up a notch on my local LLM workflow which I finally got working and gotten usable responses using the llama3:latest model from Ollama. For some reason, llama3:text was giving me some loopy responses. While looking for a drop-in upgrade I stumbled upon llama3-gradient and took it for the usual spin with the usual test:

And then, I tried to use the llama3-gradient:instruct model and it promptly lost its mind. When I say Llama 3 is chatty, this is just crazy and spat out 7+ pages. See, here:

Look, I know the promise of 8k context was that it would be quite useful as a chatbot in situations:

However, I'm a bit concerned that this type of instruct is going to quickly bore my chatbot user. But, I'm going to use it anyway and report here what happens with a few simple vs complex questions.

See ya later.