ADS 2.15.2

• Forecast Operator enhancements:
  • Fixed ETS and Theta model parameter serialization to avoid crashes when generating model parameter output.
  • Added what-if analysis support for auto-select-series workflows.
  • Updated auto-select defaults to exclude AutoMLX and AutoTS unless users explicitly select them.
  • Improved forecasting model behavior and usage guidance, including auto-select-series selection strategy documentation.
• Regression Operator enhancements:
  • Added SMAPE as a supported error metric.

ADS 2.15.1

Description

• Operators enhancements:
  • Added a new low-code Regression Operator for supervised tabular regression, with support for auto, linear_regression, random_forest, knn, and xgboost models.
  • Regression Operator runs can generate predictions, metrics, explanations, HTML reports, serialized model artifacts, and optional OCI Model Catalog and Model Deployment artifacts.
  • Forecast Operator added auto-select-series to choose a forecasting framework per series using meta-features and a meta-learning selector.
• Packaging enhancements:
  • Added the regression optional dependency extra for installing the Regression Operator dependencies.

ADS 2.15.0

Description

• Model deployment input and artifact handling enhancements:
  • ADS-generated scoring artifacts now use JSON-compatible request input handling by default for framework and HuggingFace pipeline templates.
  • Cloudpickle remains supported for trusted model artifact save/load workflows. ADS-generated scoring artifacts and ADS model input deserializers no longer deserialize cloudpickle request payloads.
  • NumPy and TensorFlow request payloads now use non-pickle NumPy buffers by default.
  • PyTorch request tensors now use structured JSON-compatible tensor payloads.
  • Existing generated artifacts loaded through ADS infer their request serializer from score.py when possible to preserve compatibility.
  • ADS-generated model artifact archives now include a file manifest. Artifact extraction and supported generated model loaders verify manifest/hash metadata when available.
• Conda pack installation enhancements:
  • ads opctl conda install now validates tar archive member paths and links before extraction.
  • Added coverage for valid archive extraction, absolute paths, path traversal, and link-target handling.
• AI Quick Actions enhancements:
  • Aqua handlers running inside Jupyter now use the standard Jupyter request preparation flow.
  • The standalone Aqua API server now uses explicit standalone and CORS settings instead of patching handler methods at runtime.
• Feature Store package refactor:
  • Removed the legacy bundled Feature Store SDK package, documentation, tests, and Feature Store marketplace operator wiring from oracle-ads.
  • Removed Feature Store specific optional dependency extras and model-save dataset linkage.
• Test improvements:
  • Updated telemetry client URL assertions to account for OCI client object-path encoding.

ADS 2.14.9

• Operators: Enhances the forecast operator
• Code refactoring

ADS 2.14.8

• Operators: Enhances the artifact copy functionality
• Code refactoring

ADS 2.14.7

• Model group: Added support for uploading artifacts for homogeneous model groups
• Forecast (what-if serve): Fixed reading model artifacts from Object Storage
• Refactor: Removed direct usage of pkg_resources

ADS 2.14.6

• AI Quick Actions enhancements
  • Removed hardcoded huggingface_hub version constraint
  • Passed MD_COMPARTMENT_OCID environment variable to the model deployment

ADS 2.14.5

• AI Quick Actions enhancements
  • Fixed Hugging Face model search: previously returned only the first 20 results; now returns up to 100.
  • Improved Shape Recommender: validates supported models and checks the structure of model configuration files.

ADS 2.14.4

• AI Quick Actions enhancements.

ADS 2.14.3

• Fixed a bug with IPython.