Thoughts on Evals(raindrop.ai)
raindrop.ai
Thoughts on Evals
https://www.raindrop.ai/blog/thoughts-on-evals/
16 comments
> Can anyone direct me towards how to ... make one?
https://hamel.dev/blog/posts/evals/
> What are "obvious" things that are important to get right - temperature set to 0? At least ~10 or 20 attempts at the same problem for each llm?
LLMs are actually pretty deterministic, so there is no need to do more than one attempt with the exact same data.
> Finally, any known/commonly used frameworks to do this, or any tooling that can call different LLMs would be enough?
https://github.com/vercel/ai
https://github.com/mattpocock/evalite
https://hamel.dev/blog/posts/evals/
> What are "obvious" things that are important to get right - temperature set to 0? At least ~10 or 20 attempts at the same problem for each llm?
LLMs are actually pretty deterministic, so there is no need to do more than one attempt with the exact same data.
> Finally, any known/commonly used frameworks to do this, or any tooling that can call different LLMs would be enough?
https://github.com/vercel/ai
https://github.com/mattpocock/evalite
"LLMs are actually pretty deterministic, so there is no need to do more than one attempt with the exact same data."
Is this true? I remember there being a randomization factor in weighing tokens to make the output more something, dont recall what
Obviously I'm not an Ai dev
Is this true? I remember there being a randomization factor in weighing tokens to make the output more something, dont recall what
Obviously I'm not an Ai dev
In my experience, the response may not be exactly the same, but the difference is negligible.
I'm very grateful! Thanks a lot
Take a look at promptfoo
Founder of a/b testing company accuses founder of evals company of misrepresenting how a/b tests are used in practice, then concludes by misrepresenting how evals are used in practice: "Or you can write 10,000,000 evals."
Could've easily been framed as "you need both evals and a/b testing," but instead they chose this route which comes across as defensive, disingenuous, and desperate.
BTW, if a competitor ever writes a whole post to refute something you barely alluded to without even mentioning their name... congratulations, you've won.
Could've easily been framed as "you need both evals and a/b testing," but instead they chose this route which comes across as defensive, disingenuous, and desperate.
BTW, if a competitor ever writes a whole post to refute something you barely alluded to without even mentioning their name... congratulations, you've won.
Agree. This whole post comes across as sales rather than the truth that both are useful for different things
The framing in this post is really weird. Automated evals can be much more informative than unit tests because the results can be much more fine grained. A/B testing in production is not suitable for determining whether all of one's internal experiments are successful or not.
I don't doubt that Raindrop's product is worthwhile to model vendors, but the post seems like its audience is C suite folks who have no clue how anything works. Do their most important customers even have any of these?
I don't doubt that Raindrop's product is worthwhile to model vendors, but the post seems like its audience is C suite folks who have no clue how anything works. Do their most important customers even have any of these?
I think in most cases, outside of pure AI providers or think AI wrappers, almost every team will realize more gains from focusing on their user domains and solving business problems versus fine tuning their prompts to eek out a 5% improvement here and there.
I don’t think you can use this as a blanket statement. For many use cases the last 5-10% is the difference between demoware and production.
If that were true, just switching to TOON would make your startup take off.
That is obviously not true because a 5% gain in LLM performance isn't going to make up for a bad product.
That is obviously not true because a 5% gain in LLM performance isn't going to make up for a bad product.
Both of these are kind of silly and vendors trying to sell you tooling you probably don't need.
In a gold rush, each is trying to sell you a different kind of shovel claiming theirs to be the best when you really should go find a geologist and and figure out where the vein is.
In a gold rush, each is trying to sell you a different kind of shovel claiming theirs to be the best when you really should go find a geologist and and figure out where the vein is.
No, thanks. Just use evals with error bars. If you can't get error bars, use an A/B test to detect spuriousness and evals.
This scared me until I realized it is about raindrop.ai, not raindrop.io.
(Raindrop.io is a bookmark service that AFAIK has "take money from people and stores their bookmarks" as its complete business model.)
(Raindrop.io is a bookmark service that AFAIK has "take money from people and stores their bookmarks" as its complete business model.)
> Intentionally or not, the word "eval" has become increasingly vague. I've seen at least 6 distinct definitions of evals
This. I am so tired of people saying evals without defining what they mean. And now even management is asking me for evals and why we are not fine tuning.
This. I am so tired of people saying evals without defining what they mean. And now even management is asking me for evals and why we are not fine tuning.
Can anyone direct me towards how to ... make one? At the most fundamental level, is it about having test questions with known, golden (verified, valid) answers, and asking different LLM models to find the answer, and comparing scores (how many were found to be correct)?
What are "obvious" things that are important to get right - temperature set to 0? At least ~10 or 20 attempts at the same problem for each llm? What are non-obvious gotchas?
Finally, any known/commonly used frameworks to do this, or any tooling that can call different LLMs would be enough?
Thanks!