Beating GPT-4 with a 13B model

[EvgeniyZh]

https://lmsys.org/blog/2023-11-14-llm-decontaminator/

[grobbyy]

I hate the clickbait title combined with the obvious result....

I have read papers like this which were good, and mostly consisted of fine-tuning, combined with a few tricks to achieve high performance in large but specific domains.

This was a waste of my life. I want the minutes of reading this and of writing this comment back.

[dchftcs → grobbyy]

It wasn't wasted, thank you for reading it so the rest of us don't.

[KennyFromIT → dchftcs]

I love this positive way of looking at it. I wish more people had this as the default - "I did a thing and it was bad/hard, here's my experience so you don't have to do the same."

[geoduck14 → grobbyy]

I would stipulate that the value wasn't "beating GPT-4" but actually "look, just because you can beat GPT-4, doesn't mean you actually delivered value" - and that is pretty important to say

[grobbyy → geoduck14]

It's possible to say that in two sentences, without a clickbait title.

[gmerc → grobbyy]

To me this reads like a cynical commentary on the state of benchmarking, not an attempt to sell a 13B GPT4-Killer

[homarp]

training on a "rephrased the test set is all you need" to ace the test set

[spoonjim]

This guy calls himself a programmer but couldn’t write code to print 100 random lines from a dictionary?

[Alifatisk → spoonjim]

Programming nowadays is slowly separating from the idea of writing code.

No-code tools has made it way easier to develope things.

Some people who barely know programming are able to create their own ai product using LangChain!