Skip to content
TheChatArchiveThe Chat Archive
Human-written demonstration (OpenAssistant volunteers)· OpenAssistant (human-written)Hugging Face OpenAssistant Conversations (oasst2)

You obviously know yourself the best, but how do you believe you learn the best? Do you…

In OpenAssistant data, “assistant” turns were written by human volunteers as demonstrations. They are not model outputs.
  1. User

    You obviously know yourself the best, but how do you believe you learn the best? Do you prefer large datasets all at once or would you rather have numerous small episodes of learning? Finally, as humans we are able to build new knowledge by leveraging what we already know, so are you able to do that easily or would you need to be re-trained/fine-tuned?

  2. Assistant

    According to latest developments, the quality of data is usually as important as quantity. Frequently learning on a small dataset of high quality specialised data can lead to better results than a large dataset of general data.

    The training is usually done as a single run, with potential rollbacks to correct loss deterioration. As long as relevant training data is saved in checkpoints, training can be split into multiple runs.

    As a large language model I can adapt to new data as long as it is present in my attention window. To gain knowledge on a permanent basis fine-tuning is required.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.