3 ms·
The current models for GPT-2 and GPT-J are large, slow and parameter-starved. Of those, the highest-performing GPT-2 models are instruction-tuned on GPT-4 and A
by smoldesu 3y ago
The current models for GPT-2 and GPT-J are large, slow and parameter-starved. Of those, the highest-performing GPT-2 models are instruction-tuned on GPT-4 and Alpaca responses.
Is it enough for "most tasks" nowadays? Maybe, but unless you're trying to sell a product to people there's really no reason not to use a LLaMA finetune.