NEWS

BERTopic 0.1.2

Fixed default safetensors/PyTorch saving on the supported Windows BERTopic 0.16.0 backend without adding an unwanted embedding-model reference.
Mapped R's pt serialization to Python's pytorch format and forwarded explicit Hugging Face model references unchanged.
Preserved original save failures instead of retrying pickle, and protected existing destinations unless overwrite is explicit.
Added regression tests for lightweight weight files, metadata restoration, reference handling, and original-error propagation.

Fixed representative-document conversion and topic-ID handling.
Validated nr_topics and synchronized returned R-side topics and probabilities after topic reduction.
Converted custom-label keys to Python integers.
Mapped document-topic matrix columns to the fitted model's actual topic IDs, including sparse output and nonconsecutive IDs.
Restored topics and probabilities when loading models and verified post-load transformation.
Made bertopic_self_check() perform and compare a deterministic fit/transform/save/load round trip.
Restored embedding models through BERTopic's backend selector.
Handled degenerate hierarchical-document visualizations and propagated visualization failures to tests.
Registered fortify.bertopic_r as a delayed ggplot2 S3 method.

Declared Python BERTopic 0.16.0 as the supported backend.
Made Conda and virtualenv installers consume one exact dependency specification and run the same import validation.
Added exact module versions to bertopic_session_info().
Added Windows Python 3.10.21 / BERTopic 0.16.0 environment and session provenance.
Added the standard testthat entry point so R CMD check runs the regression suite.
Added a Windows fresh-process benchmark harness with immutable input hashes, separate R and Python workers, version gates, raw outputs, fit/cold timing, and peak-memory recording.