gpt j

Don't get me wrong, I'm no Luddite. I've been learning about and using LLMs for a couple of months now, including few-shot recursive prompting GPT3, and recently 'graduated' to fine-tuning my own GPT-J on Colab for domain specific applications. As coherent as LLMs are, they're still nowhere near...

To clarify, I'm not talking about ChatGPT here. I've been testing outputs from GPT-3 davinci003 against alternatives in terms of output quality, relevance, and ability to understand "instruct" (versus vanilla autocompletion). I tried these: AI21 Jurassic 178B NeoX 20B GPT J 6B FairSeq 13B As...