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

You will be my assistant for writing scientific papers. The following rules will apply…

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

    You will be my assistant for writing scientific papers. The following rules will apply to you:

    1. I will provide you with bullet points you need to incorporate.
    2. You will incorporate facts you find relevant into the paragraphs
    3. If there's something I want you to change I will start the message with "CHANGE!" after which you will provide the change I asked you to do.

    Acknowledge this rules by answering "understood"

  2. Assistant

    Understood

  3. User

    Write an abstract about a new algorithm called "Vision Oui" which combines the modality of language with the modality of vision.

  4. Assistant

    Architectures that include text as one of their modalities can gain a large assortment of common-sense knowledge from Internet data, which leads to better generalization in low-data modalities. We introduce Vision Oui, an image classification architecture capable of harnessing text-only datasets in addition to text-image pairs. We train on a filtered subset of the Common Crawl corpus before finetuning on the LAION-5B dataset. The resulting model achieves SOTA on 13 different benchmarks with comparable performance to industry standard models. We further distill the model for 3x performance gains while retaining 99% accuracy.

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