#06-technical-discussion
Title
# 06-technical-discussion
v
Vishwanath Seshagiri
04/10/2023, 12:55 PMPlaying around with Codeium, Tabby, Dolly over the weekend I feel that is the way a lot of folks would go in the future, if they have the time and resources to train networks.
Many companies already have internal Analytics teams, even those that aren’t traditionally tech companies. They would be vary of sharing data with OpenAI or MS given that they don’t even use the Azure, AWS AI endpoints.
It doesn’t take a lot to train the models since they’re built to be trained with minimal data. Plus, this would allow them to re-train the model based on internal tools and libraries, instead of using LLamaIndex or some other indexing service.
👀 4
c
Conner Swann
04/10/2023, 2:39 PMI was going to dive into this later this week!
Any learnings worth sharing off the top around training?
v
Vishwanath Seshagiri
04/10/2023, 3:01 PMFor Dolly, I mostly used this tutorial:
https://github.com/sinanuozdemir/oreilly-gpt-hands-on-nlg/blob/main/notebooks/Dolly_Lite.ipynb
It was really succinct since the example is sort of contained within itself. Using a custom dataset with similar parameters would be a bit hard, and it took a lot of time to get the structure of data right, which is typically the problem in using models like this.
One thing I noticed was that Dolly had to be continuously trained on new information (similar to how RLHF works in ChatGPT).
https://github.com/TabbyML/tabby
Tabby was way more buggier than I imagined and definitely not out for consumption yet.
c
Conner Swann
04/10/2023, 3:35 PMOne thing I noticed was that Dolly had to be continuously trained on new information (similar to how RLHF works in ChatGPT).Can you elaborate on how you discovered this?
v
Vishwanath Seshagiri
04/10/2023, 3:46 PMLets take the example that is present in the document:
Copy code
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
### Instruction:
Identify the odd one out.
### Input:
Twitter, Instagram, Telegram
### Response:
TelegramOne can construe that Slack is for work, but other 3 are social apps, but Dolly or any other LLM can't figure this out like what humans can
[This is obviously ignoring the corner case that Slack can be used for social reasons]
However, Dolly would still respond with Telegram because of the way it is trained and the underlying piece of data
With RLHF (that OpenAI uses), at 10000 feet, it adapts to the data that is being given as input, thus not requiring explicit re-training for specific things. For eg, it will still return that One Medical is a publicly traded company, even though it has been acquired by Amazon.
This requires explicit retraining on new piece of information.
c
Conner Swann
04/10/2023, 3:52 PMGot it makes sense
I am thinking about this for a classification use-case, and my understanding based on what you’re saying is that if the things we’re classifying on don’t necessarily change that often, streaming fine-tuning might not be a huge deal?
v
Vishwanath Seshagiri
04/10/2023, 3:54 PMYeah, it might not be, but as is the case with anything in LLM space, it depends on the nature of information that is being trained on.
image.png
The same information of "who owns one medical" returns a more accurate answer on Bard
c
Conner Swann
04/10/2023, 3:55 PMAwesome this is neat exploration you did
thanks for sharing!!
v
Vishwanath Seshagiri
04/10/2023, 3:56 PMWith GPT-4 like models, one thing we cannot control is the underlying training data, which can have inaccuracies and biases when looking at it in context of your organization
However, Dolly has its use case when you completely trust that data sources that you have, which is true for any organization with a considerable size. That would cover almost all of the F 2000 companies, and maybe more.
c
Conner Swann
04/10/2023, 4:07 PMAs is my case indeed
o
okgodoit
04/11/2023, 2:23 AMToday I learned that one medical was acquired by Amazon. Jesus, there is no way I’m comfortable with Amazon having all my medical data 😣
7 Views