search

embeddings-benchmark / mteb

MTEB: Massive Text Embedding Benchmark
https://arxiv.org/abs/2210.07316
Apache License 2.0
1.95k stars 272 forks source link

Add proper metadata to models #1234

Open x-tabdeveloping opened 1 month ago

x-tabdeveloping commented 1 month ago

Currently, many of the fields in model metadata are not filled. These should either be specified manually or we should find ways to derive them (from HuggingFace Hub for example.)

x-tabdeveloping commented 1 month ago

We're also missing Metadata objects completely for some models (all-MiniLM-L12-v2 for instance). Or some of them have no matching revision in the results.

KennethEnevoldsen commented 1 month ago

We're also missing Metadata objects completely for some models (all-MiniLM-L12-v2 for instance). Or some of them have no matching revision in the results.

Might be worth changing the default revision to the one in the results.

x-tabdeveloping commented 1 month ago

We're also missing links

isaac-chung commented 4 weeks ago

For some of the model configs such as n_parameters, embed_dim, and max_tokens, a script can be created to get those info from HF hub like how we do it in the leaderboard code.

We're also missing links

What do you mean here?

x-tabdeveloping commented 4 weeks ago

We're also missing links

On the older leaderboard you could click on the models and it would lead you to their home page. We don't have links in the ModelMeta object right now.

For some of the model configs such as n_parameters, embed_dim, and max_tokens, a script can be created to get those info from HF hub like how we do it in the leaderboard code.

I looked into this earlier, but it's not my preferred solution. A couple of reasons:

  1. We would rely on HuggingFace Hub's interfaces and implementation.
  2. Consequently, it would be a piece of code that has to be written and maintained.
  3. It would not work for models that are not on HuggingFace Hub
  4. It would be awkward to work with models that are on the hub, but are not open access (like Llama).

I would prefer just writing these properties in by hand as:

  1. We (the team) would only have to do this once for the models we have.
  2. I think it would be relatively easy for people submitting new models as PRs to have to fill in this information.

I would love to hear others' thoughts tho, this is a slight preference, not a strong one.