OpenAssistant has been collecting instruction/response data. They’ve already used that data to refine several llama models with good success. You can also boot

OpenAssistant has been collecting instruction/response data. They’ve already used that data to refine several llama models with good success. You can also bootstrap RLHF training data from the gpt4 api. Vicuna is probably the best public model created with gpt4 data available as of today.