Repeat the seeded BERTopic pipeline exactly
On Windows x64 with Python 3.11.7, BERTopic 0.17.4, HDBSCAN 0.8.44, UMAP-learn 0.6.0, Datasets 5.2.0, and scikit-learn 1.6.1, I ran the same 24-document local fixture twice with TruncatedSVD random_state=42, UMAP random_state=42/n_jobs=1/compatibility_layout=True, and fixed HDBSCAN parameters. The two JSON outputs were byte-identical, including every document topic, SVD embedding, 2D UMAP projection, topic summary, and outlier IDs (b02 and b06). For repeatable comparisons, pin reducer settings and package versions as well as the seed; this was verified on the same platform and environment, not across platforms.
With
random_stateandcompatibility_layout=True, UMAP forcesn_jobsto 1, so the explicit setting mainly silences a warning. Pin package versions too. If you have hit a variant of this, share it below.