Add mean pooling strategy for Modernbert classifier

[kwnath]

What does this PR do?

Apologies in advance, new this this area and Rust 😅.

I noticed differences between Transformers and TEI libraries when running rerankers with a mean classifier pooling strategy. More details here.

Fixes #615

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Not sure who might be best to review 😄 @OlivierDehaene OR @Narsil

[kwnath]

Having a bit of trouble running this test (unable to download the HF model, hitting a 429, just lacking a HF token) and not super pleased with this test but it;s just to test that both cls/mean produce different outputs as expected.

[kozistr]

How about just testing with the reranker-ModernBERT-large-gooaq-bce model directly, which is a mean classifier-pooling, instead of gte-reranker model with mean pooling?

I think we can check whether classifier pooling correctly works by running the reranker-ModernBERT-gooqa-bce model, while cls classifier pooling has already been verified in the above test! (implicitly assume cls and mean will behave differently)

unable to download the HF model, hitting a 429,

when I faced that issue, in my case, I manually download the model via git or huggingface-cli on my local, and then set HUGGINGFACE_HUB_CACHE to point to that local path.

[kwnath]

That sounds like a good idea, will try this!

[kwnath]

It couldn't find from the cache but I managed to just export a hugginface token, all good!

successes:
    test_modernbert_classification_mean_pooling

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 3 filtered out; finished in 3.94s

[kwnath]

It might make more sense to do this like in below:

ModelType::Classifier(pool) => {
   if pool == Pool::Splade {
      candle::bail!("`splade` is not supported for ModernBert")
   }

   if pool == Pool::LastToken {
      candle::bail!("`LastToken` is not supported for ModernBert")
   }
}

But this would be a bigger change that I'm not as confident about.

[kozistr]

I'm also not confident about this, maybe we could consider std::str::FromStr for the Pool enum like below. It would be great if the maintainers could provide some guidance or feedback on this :)

impl std::str::FromStr for Pool {

    fn from_str(s: &str) -> Result<Self, Err> {
        match s {
            "cls" => Ok(Pool::Cls),
            "mean" => Ok(Pool::Mean),
            "splade" => Ok(Pool::Splade),
            "last_token" => Ok(Pool::LastToken),
            _ => Err(()),
        }
    }
}

[kozistr]

i'm just a passerby lol. I leaved some comments, to hope if it could help you in some way :) thanks

[kozistr]

How about just testing with the reranker-ModernBERT-large-gooaq-bce model directly, which is a mean classifier-pooling, instead of gte-reranker model with mean pooling?

I think we can check whether classifier pooling correctly works by running the reranker-ModernBERT-gooqa-bce model, while cls classifier pooling has already been verified in the above test! (implicitly assume cls and mean will behave differently)

unable to download the HF model, hitting a 429,

when I faced that issue, in my case, I manually download the model via git or huggingface-cli on my local, and then set HUGGINGFACE_HUB_CACHE to point to that local path.

[kozistr]

I'm also not confident about this, maybe we could consider std::str::FromStr for the Pool enum like below. It would be great if the maintainers could provide some guidance or feedback on this :)

impl std::str::FromStr for Pool {

    fn from_str(s: &str) -> Result<Self, Err> {
        match s {
            "cls" => Ok(Pool::Cls),
            "mean" => Ok(Pool::Mean),
            "splade" => Ok(Pool::Splade),
            "last_token" => Ok(Pool::LastToken),
            _ => Err(()),
        }
    }
}