top | item 44852533

LLM Evals Are Just Tests. Why Are We Making This So Complicated?

3 points| camwest | 6 months ago |cameronwestland.com

2 comments

order

8organicbits|6 months ago

So, did the tests allow you to build a system that never confused existing features with new features? That seems like the problem statement, but I think I'm only seeing probabilistic testing.

camwest|6 months ago

Never? No. Way less likely? Yes!

In dev we do 100 consistency checks and get green. In CI we do 10.