3.0.0: Improved ranking

Summary

In this release, we added a return_ranks and a rank column to some UDFs,
to improve the flexibility in building Pipelines using the Exasol Transformers Extension.

BREAKING CHANGES:

• The Sequence-Classification UDF's (single text, text pair) now require a return_ranks input Column, which is used
  to determine how many results per input should be returned.
• The Sequence-Classification UDF's (single text, text pair) now return a rank column with the rank of the results.
• The Zero-Shot-Classification UDF now require a return_ranks input Column, which is used to determine how many
  results per input should be returned.

Please refer to the user guide for Details.

Features

• #329: Added return_ranks to Zero-Shot-Classification UDF
• #326: Added parameters rank and return_ranks to sequence classification single text udf
• #327: Added parameters rank and return_ranks to sequence classification text pair udf

Refactorings

• #311: Updated transitive dependencies
• #337: Moved CI Integration tests from AWS CodeBuild to GitHub Actions

Security

• #333: Resolved CVE-2025-3730 for torch by bumping version to ^2.8.0 and setting requires-python = ">=3.10.0,<3.15"

Internal

• #333: Resolved CVE-2025-8869 for transitive dependency pip by re-locking dependencies

2.3.0

Transformers Extension 2.3.0, 2025-09-24

Code name: Model utility UDF's

Summary

In this release we added utility functionality for installing,
listing and deleting HuggingFace models in the Exasol BucketFS.

Features

• #149: Added Python API for uploading a model to a given BucketFS location
• #138: Added new UDF to delete an uploaded model
• #313: Added new UDF to list installed models in a BucketFS directory

Bugfixes

• #285: Fixed logic in model output quality checks of UDF integration test
• #285: Deactivated SaaS tests because of a broken API

Documentation

• #319: Updated TE User Guide

Refactorings

• Updated tornado (6.4.2 -> 6.5.1)
• #201: Added python toolbox to project
• #294: Improved linter score
• #305: Updated python toolbox to 1.1.0 & added basic typing checks
• #295: Activated type check in .pre-commit-config.yaml
• #296: Activated code formatting in .pre-commit-config.yaml
• #311: Updated transitive dependencies urllib3 (2.4.0 -> 2.5.0) & requests (2.32.3 -> 2.32.4)
• #317: Added functions install_huggingface_model() and load_huggingface_pipeline().

2.2.1: Minor Update for internal changes

Summary

Fixed a bug in a model fixture, and updated exasol-bucketfs to 1.0.1

Bugs

• #287: Used the correct AutoModel class in tests.fixtures.model_fixture_utils.download_model_to_path

Dependency Updates

• Updated exasol-bucketfs to 1.0.1

2.2.0: Bugfix for token classification

Summary

This release includes a bugfix for handling unexpected results in the token classification udf,
as well as internal refactorings for the unit tests.

Bugs

• #272: Fixed unit tests assertions not working correctly
• #275: Fixed a bug where models returning unexpected results was not handled correctly

Refactorings

• #273: Refactored unit tests for token_classification_udf to use StandAloneUDFMock, made params files more maintainable
• #271: Moved test cases which only pertain to the base udf to base udf unit tests

2.1.0: Added new deploy command

Summary

This release features a new CLI command for deploying the Transformers Extension,
additionally improvements where made for the tests.

Bugs

• #256: Removed imports of pytest fixtures to avoid session fixtures running for each test

Refactorings

• #252: Use the pytest plugins for in the integration tests.
• #264: Created a single deployment CLI command.

2.0.0: Fixed model saving, added SaaS support and update to Python 3.10

Summary

This release Fixes an error in saving and loading of the model metadata. It also adds Exasol Saas support and
updated the project to python 3.10

Features

• #243: Added an option to deploy scripts in a SaaS database.
• #244: Made the integration tests running in SaaS, as well as in the Docker-DB.

Bugs

• #237: Fixed reference to python-extension-common
• #245: Added task_type parameter to fix model saving and loading

Documentation

• #210: Fixed typos in user guide.
• #247: Updated documentation including the deployment options in SaaS.

Refactorings

• #216: Simplified model path constructions, consolidating them into one function
• #228: Now use python-extension-common for the language container deployment.
• #232: Added Class which holds model information
• #217: Refactored PredictionUDFs and LoadLocalModel so that LoadLocalModel constructs the bucketfs model file path
• #230: Updated supported python version to >= Python 3.10
• #236: Moved to the PathLike bucketfs interface.
• #218: Changed upload_model_udf to load model from Huggingface

Security

1.0.1: Fixed the directory structure bug

Summary

Attention: In the 1.0.x Releases the model saving process is broken. Some metadata is not saved correctly resulting in executable models which return bad results. If you use locally saved models that you upload manually you are not affected, otherwise, please skip this version and wait for release 2.0.0. We are working on fixing this issue.

Fixed the directory structure made by the model upload UDF.

Bugs

• #221: Directory Structure that Model Upload UDF creates is different from what PredictionUDFs expect.

Features

N/A

Refactorings

N/A

Security

N/A

1.0.0: Local model loading

Summary

Attention: In the 1.0.x Releases the model saving process is broken. Some metadata is not saved correctly resulting in executable models which return bad results. If you use locally saved models that you upload manually you are not affected, otherwise, please skip this version and wait for release 2.0.0. We are working on fixing this issue.

In this release, we integrated a new model loading functionality which means downloaded models will now be saved
in the BucketFS. This means, the Prediction UDFs do not connect to the internet to look for model updates.
There are also documentation updates, and we updated cryptography to >= 42.0.4.

Breaking API changes

The change in the model loading functionality means the API for the Prediction UDFs has changed.
The 'token_conn' parameter was removed from the UDF calls. You can now call the UDFs
as follows (Example case for the filling mask udf):

SELECT TE_FILLING_MASK_UDF(
    device_id,
    bucketfs_conn,
    sub_dir,
    model_name,
    text_data,
    top_k
)

Features

• #205: Added vagrant setup
• #146: Integrated new download and load functions using save_pretrained

Documentation

• #133: Improved user and developer documentation with additional information

Refactorings

• #147: Removed token_conn from Prediction UDFs

Security

• Updated cryptography to >= 42.0.4

0.10.0: Fixing deployment issues.

Summary

Deploying SLC under Windows, releasing to PyPi.

Features

• #185: Release to pypi during release workflow

Bug Fixes

• #166: Error deploying language container under Windows

Refactorings

Security

0.9.2: Updated and Fixed Dependencies III

Summary

Changed the dependency on exasol-bucketfs and typeguard.

Features

N/A

Refactorings

• #192 (revisited): Changed the dependency on exasol-bucketfs

Security

N/A