An LLM fine-tuning course online conference for everything LLMs.
Build skills to be effective with LLMs
Course website: https://maven.com/parlance-labs/fine-tuning
<<< Fine-Tuning Workshop 2 | Fine-Tuning Workshop 3 >>>
Fine tuning when you've already deployed LLMs in production.
Date: May 23, 2024
Time: 11:00 PM - 12:00 PM GMT
Speaker: Kyle Corbitt (Twitter)
MC: Hugo Bowne-Anderson (Twitter)
Already have a working prompt deployed in production? Fine-tuning may be significantly easier for you, since you’re already collecting training data from your true input distribution! We’ll talk through whether it’s a good idea to replace your prompt with a fine-tuned model at all, and the flow we’ve found most effective if you choose to do so. We’ll also review important gotchas to watch out for like data drift post-deployment.
(WIP)
Popcorn kernel jokes - funny how that quickly turned into baldness (dad jokes?). Haha
Kyle Corbitt's background:
Before founding OpenPipe, he's done a lot of things. He led the Startup School team at Y Combinator, which was responsible for the product and content that YC produces for early stage companies. Prior to that he worked in engineering at Google and studied machine learning at school.
How many LLM calls you make in your life?
Kyle: I'm really loving this, this beautiful normal distribution we've got on the poll question.
Poll results:
We're at an 80% response rate.
This is interesting to me. We've got this nice, normal-ish looking curve here. About half of the audience is 10,000 or less, and then half is more, including 6 folks who've done at least a million. So it sounds like the 6 folks definitely have LLMs in production.
If we look at everyone with a 1,000 or more calls, that's about 25 out of the audience. Then it looks like comfortably the "fat" part of the curve there's less than 1,000 calls or may be not currently using LLMs in production.
All which is actually great.
You know, I know the title is about what to do if you're already doing in production. But actually, I think I've got quite a lot I want to talk about on how to get that process started as well, or how to think about it at least while you're starting.
(watch the video at 00:12:11)
(watch the video at 00:13:51)
(WIP)
(watch the video at 00:42:09)
(WIP)
(watch the video at 00:44:51)
(WIP)
That's pretty much it.
Slide:
- $222.22 in OpenPipe credits!
- Email: [email protected]
- "FT Course credits" in the subject line
(watch the video at 00:45:44)
(WIP)
Questions?
(watch the video at 00:46:17)
(WIP)
(WIP)
My collected links. Not sure if it's complete, and also it's not that much this time:
- https://openpipe.ai/
- https://hamel.dev/notes/llm/inference/03_inference.html
- Argilla - Argilla is an open-source data curation platform for LLMs. We provide support for each step in the MLOps cycle. Kyle uses it for data labeling.
- https://docs.argilla.io/en/v1.1.0/guides/steps/1_labelling.html
- https://huggingface.co/spaces/mteb/leaderboard
- https://www.answer.ai/posts/2024-04-26-fsdp-qdora-llama3.html
- https://twitter.com/corbtt](https://twitter.com/corbtt
Source: Discord (thanks to CodingWitcher)
Some highlights:
(WIP)