8.12.0

Added

• Supported XGBRanker model (#649)
• Accepted LTR (Learning to rank) model config when importing model (#645, #651)
• Added LTR feature logger (#648)
• Added prefix_string config option to the import model hub script (#642)
• Made online retail analysis notebook runnable in Colab (#641)
• Added new movie dataset to the tests (#646)

8.11.1

Added

• Make demo notebook runnable in Colab (#630)

Changed

• Bump Shap version to 0.43 (#636)

Fixed

• Fix failed import of Sentence Transformer RoBERTa models (#637)

8.11.0

Added

• Support E5 small multilingual model (#625)

Changed

• Stream writes in ed.DataFrame.to_csv() (#579)
• Improve memory estimation for NLP models (#568)

Fixed

• Fixed deprecations in preparation of Pandas 2.0 support (#602, #603, contributed by @bartbroere)

8.10.1

Fixed

• Fixed direct usage of TransformerModel (#619)

8.10.0

Added

• Published pre-built Docker images to docker.elastic.co/eland/eland (#613)
• Allowed importing private HuggingFace models (#608)
• Added Apple Silicon (arm64) support to Docker image (#615)
• Allowed importing some DPR models like ance-dpr-context-multi (#573)
• Allowed using the Pandas API without monitoring/main permissions (#581)

Changed

• Updated Docker image to Debian 12 Bookworm (#613)
• Reduced Docker image size by not installing unused PyTorch GPU support on amd64 (#615)
• Reduced model chunk size to 1MB (#605)

Fixed

• Fixed deprecations in preparation of Pandas 2.0 support (#593, #596)

Release 8.9.0

Added

• Simplify embedding model support and loading #569
• Make eland_import_hub_model easier to find on Windows #559
• Update trained model inference endpoint #556
• Add BertJapaneseTokenizer support with bert_ja tokenization configuration #534
• Add ability to upload xlm-roberta tokenized models #518
• Tolerate different model output formats when measuring embedding size #535
• Generate valid NLP model id from file path #541
• Upgrade torch to 1.13.1 and check the cluster version before uploading a NLP model #522
• Set embedding_size config parameter for Text Embedding models #532
• Add support for the pass_through task #526

Fixed

• Fixed black to comply with the code style #557
• Fixed No module named 'torch' #553
• Fix autosummary directive by removing hack autosummaries #548
• Prevent TypeError with None check #525

8.7.0

Added

• Added a new NLP model task type "text_similarity" (#486)
• Added a new NLP model task type "text_expansion" (#520)
• Added support for exporting an Elastic ML model as a scikit-learn pipeline via MLModel.export_model() (#509)

Fixed

• Fixed an issue that occurred when LightGBM was installed but libomp wasn't installed on the system. (#499)

8.3.0

Added

• Added a new NLP model task type "auto" which infers the task type based on model configuration and architecture (#475)

Changed

• Changed required version of 'torch' package to >=1.11.0,<1.12 to match required PyTorch version for Elasticsearch 8.3 (was >=1.9.0,<2) (#479)
• Changed the default value of the --task-type parameter for the eland_import_hub_model CLI to be "auto" (#475)

Fixed

• Fixed decision tree classifier serialization to account for probabilities (#465)
• Fixed PyTorch model quantization (#472)

8.2.0

Added

• Added support for passing Cloud ID via --cloud-id to eland_import_hub_model CLI tool (#462)
• Added support for authenticating via --es-username, --es-password, and --es-api-key to the eland_import_hub_model CLI tool (#461)
• Added support for XGBoost 1.6 (#458)
• Added support for question_answering NLP tasks (#457)

8.1.0

Added

• Added support for eland.Series.unique() (#448, contributed by @V1NAY8)
• Added --ca-certs and --insecure options to eland_import_hub_model for configuring TLS (#441)