Transformers Extension AI Functions (Release 5.2.0)

Product News
, ,
1

We released version 5.2.0 of the Transformers Extension on Github and Pypi.
The Exasol Transformers Extension allows you to use pre-trained machine learning models from :hugging_face: Hugging Face directly in your Exasol instance.
It lets you install the models via the transformers-api directly into Exasol’s filesystem BucketFS, and use them on your data using provided UDF’s.

:memo: Summary

Since our last announcement, our AI Functions set has been completed. AI_ANSWER, AI_CLASSIFY, AI_SENTIMENT, AI_TRANSLATE and AI_EXTRACT_ENTITIES are now available.
Additionally, the upload of models to the BucketFS has gotten a rework to allow the use of bigger models.
We also added a new script which installs all available UDF’s in Database.
Lastly, the transformers package was updated from version 4 to version 5.

:warning: BREAKING CHANGES:

In transformers version 5, the pipeline for question-answering has been removed. For our AI_ANSWER_EXTEDNED UDF, we replaced it with the text-generation pipeline. You will not be able to use you old question-answering models with this new implementation, switch to text-generation models instead. Make sure these models support using a chat_template(instruct models).

(The translation pipeline was also removed, but we decided to maintain it ourselves for
now, so the AI_TRANSLATE_EXTENDED UDF is unaffected.)

1 Like
2

Thanks for the release note, Marlene. The completed AI Functions set is a nice milestone, especially having translate, sentiment, classification and entity extraction usable directly in Exasol without wiring half a zoo around it. The bigger model upload rework also sounds very welcome, BucketFS model handling was usually where reality started throwing chairs. Small heads-up: there’s a typo in “AI_ANSWER_EXTEDNED”, unless the function name really contains that spelling. Also good call on calling out the question-answering pipeline change clearly, that one can bite people hard during upgrades.