underrated use case of chatgpt's new memory feature 😉 i said 'no yapping; remember this'

Shreya has been a delight to work with! If you are doing anything related to LLMOps, reach out to her!🙂

got screwed by my evals today. lesson: if you rely on a separate labeled dataset for testing changes to your production LLM pipeline, and some prod LLM inputs may consist of errors like typos, bad grammar, different capitalization, etc. the test set needs to reflect this!!

underappreciated in LLM-land is how humans ultimately decide what matters, what is valuable, and what is right.

Who validates the validators?

An interesting observation from this paper is even Humans are not reliable judges of LLMs, because their evaluation criteria drift as they spend more time with LLMs.

Maybe the approach proposed in the paper can be incorporated into chatbot arena.

👩‍🏫 Who validates the validators?

Shreya Shankar et al. introduced EvalGen, a mixed-initiative approach that aligns LLM-generated evaluations of LLM outputs with human preferences, addressing inherent biases and improving reliability.

Implementation details:
arxiv.org/abs/2404.12272

The Multi-Eval node is now in ChainForge!🎉 Define multiple criteria and evaluators on one node, including mix of code- or LLM-based evals. Table View is also improved.📈 Checkout v0.3.1.5 on PyPI, or on chainforge.ai/play. Release notes follow.

We'll roll out EvalGen in ChainForge in the coming weeks. This week, I aim to push the Multi-Eval node, alongside the nifty improvements to the table view for showing scores across many evaluators.

100% relate to this. one reason that evaluation criteria is hard to work out in advance is that you don't know ahead of time how the model's going to fail, and that's usually what you're looking for in evals

ie. there's no need to do evals over something that you know will…

Shreya Shankar This is a super interesting workflow and something we are doing really manually currently. Build evals > test using sample with human GT labels > refine eval. Then, in dev, run evals routinely, look for increased scores, but review outputs to apply qualitative adjustment.

Evals are arguably the hardest part of LLMOps. LLMs mess up, so we check them w/ other LLMs, but this feels icky. Who validates the validators??

We built an interface to align LLM-based evals with user preferences, learning a lot about why this is hard: arxiv.org/abs/2404.12272

This has been one of the things I’ve been most concerned about when it comes to LLMs.

Scaling user judgements is hard so I’m looking forward to reading into this 👀

since I'm deeply immersed in evals right now (and the process of building them) I got a kick out of this paper from Shreya Shankar J.D. Zamfirescu bjoern hartmann @[email protected] Aditya Parameswaran Ian Arawjo (@[email protected])

it addresses the challenge of time-efficiently coming up with evals that are aligned with practitioners

some…