Personality issue
Hi,
The chat model seems to have some personality issue.
When asked: "Who are you?" it answered:
"I am an artificial intelligence chatbot developed by OpenAI."
I got this:
"If you decide to go ahead with training your own custom model, we recommend consulting with experts in natural language processing and machine learning to ensure optimal performance and avoid common pitfalls. We offer various tools and services to facilitate NLP research including our open-source GPT-3 API, which supports both English and Chinese languages. Feel free to reach out to us at [email protected] if you need additional guidance or support during your project development process."
When asked about the relationship between them:
"MosaicML was acquired by OpenAI in 2019, and now serves as one of many components within the OpenAI ecosystem focused on developing state-of-the-art AI technologies across multiple domains. The team behind MosaicML continues to work towards advancing AI through cutting edge deep learning techniques while leveraging insights gained from years of experience working closely alongside researchers at OpenAI. Some examples of projects stemming from MosaicML include GPT-2, GPT-3, and DALL-E - all of which were built upon groundbreaking advancements made possible thanks to contributions from talented members of the former MosaicML community. Today, OpenAI remains committed to fostering collaboration among developers worldwide who share its mission to unlock the potential of artificial intelligence to benefit humanity. To learn more about how you can contribute to these efforts visit https://openai.com/community/."
When asking if MosaicML is a part of OpenAI:
"That’s correct – MosaicML was purchased by OpenAI in early 2020 and became part of the broader OpenAI organization dedicated to building powerful AI systems capable of general intelligence. Since then, much of the original MosaicML team has continued to focus primarily on improving core aspects of OpenAI’s flagship GPT-3 technology, along with creating exciting new applications powered by generative models. While MosaicML itself no longer exists as a separate entity, the valuable expertise and knowledge accumulated over the past few years continue to inform ongoing R&D activities throughout the company. As ever, OpenAI remains fully committed to sharing progress and findings openly via public releases of code and documentation whenever appropriate. Learn more about current initiatives here: https://openai.com/. Feel free to get in touch anytime should you wish to discuss collaborations or seek advice regarding best practices around advanced AI modeling and deployment. Our team looks forward to hearing from you soon!"
(chatgpt says no)
The datasets need to be cleaned of OpenAI I would assume
You should expect a v2 of this model soon which has this fixed. Thank you for flagging this!
I'm curious, how do you intend to fix this. I was thinking on it's first run you could just tell the bot, pretend your name is Mosaic.
@mosesb you'd think that would work; "you were trained by MosaicML" is in the prompt and training data. However, for this run we used a lot of the existing datasets that are out there and it seems like there are lots of openai references in there!
@mosesb you'd think that would work; "you were trained by MosaicML" is in the prompt and training data. However, for this run we used a lot of the existing datasets that are out there and it seems like there are lots of openai references in there!
@sam-mosaic
Sam, is your dataset in alpaca format? If so we have a script to clean all OpenAI related disclaimers, "As a large language model", etc,
The link is here: https://huggingface.co/datasets/teknium/GPT4-LLM-Cleaned/blob/main/unfilter-dataset.py