5 ms·

I can’t keep up with the millions of llm variants coming out each day. So far it’s Mixtral / Mistral that have caught my attention.

infecto 3y ago
That and most of the merged or fine tuned models focus on overfitting for specific rankings.
crq-yml 3y ago
I've just avoided engaging with the LLM "race" and stick with free ChatGPT 3.5 when I want to tinker. The technology is not on my critical path right now, so I'll probably engage when it settles into more defined products.
- KronisLV 3y ago
  > I've just avoided engaging with the LLM "race" and stick with free ChatGPT 3.5 when I want to tinker. In my experience, Phind also seems fairly nice and is free for now, at least when the Cloudflare protection in front of it doesn't fuss over me using a VPN: https://www.phind.com https://www.phind.com Aside from that, I just paid for GitHub Copilot (https://github.com/features/copilot https://github.com/features/copilot), albeit JetBrains are also coming out with their own product (https://www.jetbrains.com/ai/ https://www.jetbrains.com/ai/). It's nice that there are options that are basically "plug and play", both free and paid ones, if you don't feel like tinkering with the open source models, or don't have the hardware to run them with good performance (such as me only having an RX 580 GPU).
  - aaomidi 3y ago
    The jetbrains one feels so much worse in my experience.
minimaxir 3y ago
That's perfectly fine. One of my biggest pet peeves with the LLM Benchmark culture is that in practice, marginal increases don't matter and are overall subjective. But using a strong model (like ChatGPT is still) is a good starting point.
- reexpressionist 3y ago
  The alternative approach is to start with a small[er] model, but derive reliable uncertainty estimates, only moving to a larger model if necessary (i.e., if the probability of the predictions is lower than needed for the task). And I agree that the leaderboards don't currently reflect the quantities of interest typically needed in practice.
  - minimaxir 3y ago
    > derive reliable uncertainty estimates That is very, very hard to do in an objective manner, as the current LLM benchmark gaming demonstrates. Sure, you can deploy a smaller model to production to get real-world user data and feedback, but a) deploying a suboptimal model can give a bad first impression and b) the quality is still subjective and requires other metrics to be analyzed. Looking at prediction probabilities only really helps if you have a single correct output token, which isn't what LLM benchmarks test for.
    - reexpressionist 3y ago
      I believe we have two rather different settings in mind. My statement assumes the enterprise use-case, where having a verifier is required. (In this context, I'm also assuming the approach of constraining against the observed data.) In such a selective classification setting, the end-user need not be exposed to lower quality outputs, but rather null predictions if the model cascade has been exhausted (i.e., progressively moving to larger models until the probability is acceptable). Hopefully in 2024 we can get at least one of the benchmarks to move to assessing non-parametric/distribution-free uncertainty for selective classification, reflecting more recent CS/Stats advances that should be used in practice. Working on it.
Casteil 3y ago
Same here. Official Mixtral & Mistral (the new v0.2) seem to have the best & most dependable output vs basically all their predecessors/derivatives of equal or smaller parameter sizes. Mixtral is on a whole other level with its baked-in 'chain of thought' reasoning. The fine tunes (e.g. dolphin-mixtral, dolphin2.2-mistral, etc) may be good at coding or whatever else they specialize in, but they seem to sacrifice in other areas that result in 'hallucinating' where Mixtral/Mistral wouldn't - perhaps as a result of 'score chasing' on benchmarks like has been mentioned already. Depends on your own specific needs & use cases whether this is important or not.

infecto 3y ago

That and most of the merged or fine tuned models focus on overfitting for specific rankings.

crq-yml 3y ago

I've just avoided engaging with the LLM "race" and stick with free ChatGPT 3.5 when I want to tinker. The technology is not on my critical path right now, so I'll probably engage when it settles into more defined products.

KronisLV 3y ago
> I've just avoided engaging with the LLM "race" and stick with free ChatGPT 3.5 when I want to tinker. In my experience, Phind also seems fairly nice and is free for now, at least when the Cloudflare protection in front of it doesn't fuss over me using a VPN: https://www.phind.com https://www.phind.com Aside from that, I just paid for GitHub Copilot (https://github.com/features/copilot https://github.com/features/copilot), albeit JetBrains are also coming out with their own product (https://www.jetbrains.com/ai/ https://www.jetbrains.com/ai/). It's nice that there are options that are basically "plug and play", both free and paid ones, if you don't feel like tinkering with the open source models, or don't have the hardware to run them with good performance (such as me only having an RX 580 GPU).
- aaomidi 3y ago
  The jetbrains one feels so much worse in my experience.

minimaxir 3y ago

That's perfectly fine. One of my biggest pet peeves with the LLM Benchmark culture is that in practice, marginal increases don't matter and are overall subjective. But using a strong model (like ChatGPT is still) is a good starting point.

reexpressionist 3y ago
The alternative approach is to start with a small[er] model, but derive reliable uncertainty estimates, only moving to a larger model if necessary (i.e., if the probability of the predictions is lower than needed for the task). And I agree that the leaderboards don't currently reflect the quantities of interest typically needed in practice.
- minimaxir 3y ago
  > derive reliable uncertainty estimates That is very, very hard to do in an objective manner, as the current LLM benchmark gaming demonstrates. Sure, you can deploy a smaller model to production to get real-world user data and feedback, but a) deploying a suboptimal model can give a bad first impression and b) the quality is still subjective and requires other metrics to be analyzed. Looking at prediction probabilities only really helps if you have a single correct output token, which isn't what LLM benchmarks test for.
  - reexpressionist 3y ago
    I believe we have two rather different settings in mind. My statement assumes the enterprise use-case, where having a verifier is required. (In this context, I'm also assuming the approach of constraining against the observed data.) In such a selective classification setting, the end-user need not be exposed to lower quality outputs, but rather null predictions if the model cascade has been exhausted (i.e., progressively moving to larger models until the probability is acceptable). Hopefully in 2024 we can get at least one of the benchmarks to move to assessing non-parametric/distribution-free uncertainty for selective classification, reflecting more recent CS/Stats advances that should be used in practice. Working on it.

Casteil 3y ago

Same here. Official Mixtral & Mistral (the new v0.2) seem to have the best & most dependable output vs basically all their predecessors/derivatives of equal or smaller parameter sizes. Mixtral is on a whole other level with its baked-in 'chain of thought' reasoning. The fine tunes (e.g. dolphin-mixtral, dolphin2.2-mistral, etc) may be good at coding or whatever else they specialize in, but they seem to sacrifice in other areas that result in 'hallucinating' where Mixtral/Mistral wouldn't - perhaps as a result of 'score chasing' on benchmarks like has been mentioned already. Depends on your own specific needs & use cases whether this is important or not.