Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[RFC] Support more local model types #1164

Open
ylwu-amzn opened this issue Jul 29, 2023 · 7 comments
Open

[RFC] Support more local model types #1164

ylwu-amzn opened this issue Jul 29, 2023 · 7 comments
Labels
RFC Request For Comments from the OpenSearch Community

Comments

@ylwu-amzn
Copy link
Collaborator

ylwu-amzn commented Jul 29, 2023

Currently, ml-commons only supports uploading text-embedding models. However, we believe there are other models that could be valuable additions to our platform:

  • Summarization model: This model can summarize a document or lengthy content and return a concise version.
  • Question answering model: If you have a question related to a context or set of documents, this model can provide you with accurate answers.
  • Text classification: This model is designed to classify text, such as sentiment analysis, where it assigns labels like 'positive,' 'negative,' or 'neutral.'
  • Named entity recognition (NER) model: For identifying various entities in a text, such as organizations, persons, locations, and more.
  • Image embedding model: This model translates an image into a vector representation for easier analysis.
  • Object detection model: This model detects object in image
  • Rerank model: rerank the search result

Please comment on this issue if you require support for other local models or vote for the model you need the most.

@ylwu-amzn ylwu-amzn added enhancement New feature or request untriaged labels Jul 29, 2023
@hijakk
Copy link

hijakk commented Jul 29, 2023

+1. Most of these would be valuable in my use cases, as well as language identification. Possibly langid is a different beast than this approach should support?

Image vectorization is a potentially tough use case as images can be large and including them in base64 natively in documents can dramatically inflate document size on disk rather than providing a reference pointer to external storage (such as s3).

@ylwu-amzn
Copy link
Collaborator Author

From #1150 (comment)

@asfoorial suggests

I suggest to keep the door open for LLM hosting as there is a trend to get LLMs smaller with quantization yet achieve reasonable performance. I would say they will be hostable in ml nodes or other dedicated nodes.

@nateynateynate
Copy link
Member

They all get a thumbs up from me, but I actually would love to see image embedding. I'm fascinated by it.

@HungryHowies
Copy link

HungryHowies commented Aug 3, 2023

+1 tbh , I would love to have them all.

@austintlee
Copy link
Collaborator

Are cross encoders covered under "rerank model"?

@austintlee
Copy link
Collaborator

@dhrubo-os (tagging you since you went over some basics on this with the OCI students)

Would it make it easier to produce ML input/output classes for all these different models if we used Smithy to define them and have it generate the classes. Just wondering what we can do to expedite progress on this using some common framework.

@TrungBui59
Copy link
Contributor

@ylwu-amzn @dhrubo-os I am interested in working on supporting the Question-answering model. Can you guys give me some hints on what I should do? Currently, I am thinking of following the approach that we used to support the text-embedding model

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
RFC Request For Comments from the OpenSearch Community
Projects
Status: On-deck
Development

No branches or pull requests

6 participants