MLflow Autologging Integration for BERTopic

[NJAHNAVI2907]

Feature request

I am currently implementing an autologging integration for BERTopic with MLflow for managing ML experiments. This would automatically log BERTopic's training parameters (e.g., embedding model, UMAP/HDBSCAN settings), metrics, artifacts , and the fitted model via MLflow's PyFunc flavor during fit_transform calls. The goal is to simplify experiment tracking in BERTopic workflows without manual logging.

Motivation

Approach: Monkey-patching BERTopic.fit_transform using MLflow's safe_patch for safe integration.
What Gets Logged:
Parameters: Embedding model name, UMAP (n_neighbors, n_components, ...), HDBSCAN (min_cluster_size...), vectorizer type
Metrics: n_documents, avg_doc_length, n_topics, n_outliers, avg/max/min_topic_size, vocab_size, embedding_dim, diversity, coherence (c_v, c_npmi, u_mass via gensim), per-topic coherence.
Artifacts: topic_info.csv, metrics.json, per_topic_coherence.csv, embeddings.npy and the full model as PyFunc.
Flavor Support: Registered as an MLflow flavor (bertopic) with @autologging_integrationenabling mlflow.autolog()

Mlflow issue : mlflow/mlflow#16792 (comment)

Your contribution

Seeking Feedback:

Is this something you would be interested in merging into BERTopic's core as an optional MLflow submodule or would it be better as an external package like mlflow-scikit-learn or mlflow-txtai.

Should I pursue adding this as a separate repo with a lazy import in MLflow's init.py via PR or integrate it directly into BERTopic? Pros/cons from your perspective?

I would like to hear your thoughts on this. Thank you!