Breaking Changes

• HDP dead topics dropped. get_phi() / get_theta() return only live topics for tomotopy HDPModel. Fewer topics than before, and columns/index now carry the model's own ids (e.g. [6, 8, 24, 25]) instead of 0..k-1. Warns when dropping. Other model types unaffected.
• bitermplus >= 1.0 required. p(z|d) is no longer stored on BTM, so get_theta() needs the documents: get_theta(model, corpus=get_vectorized_docs(docs, vocab)). BTMClassifier needs no corpus.
• entropy() raises for a single topic. Rényi entropy is undefined at T=1 (q = 1/T = 1 zeroes the F/(q-1) denominator). Previously returned an arbitrary fallback. Matches bitermplus 1.0.
• get_closest_topics() rejects out-of-range or negative ref instead of silently clamping; empty model list errors.
• get_stable_topics() validates array shapes and ref.

New Features

• entropy() accepts either orientation. W×T (as get_phi returns) or T×W, inferred from which axis sums to 1. New topics_axis argument states it explicitly; a value contradicting the data now errors instead of returning a wrong number.
• BTMClassifier support, end to end — get_phi, get_theta, prepare_coords, report.
• prepare_coords() accepts precomputed phi / theta.

Bug Fixes

• entropy() silently returned wrong numbers for a transposed matrix — 1.1625 vs the correct 0.7272 on one 40×30 phi, with no error. It selects the topic count, so this could change published conclusions.
• Dead HDP topics corrupted everything downstream — entropy 0.0995 where the true value was 1.5524.
• lle / ltsa crashed on collapsed topic distances ("output dimension must be less than or equal to input dimension").
• An HDP model with no live topics returned an empty phi; now raises.
• Jaccard top_words=0 raised ZeroDivisionError; negative values silently returned wrong distances.
• report() failed when docs was a NumPy array.
• Selected topic was not highlighted when using a custom topic_col.
• Duplicate word labels pushed genuine terms out of calc_terms_probs_ratio().
• get_top_docs() gave an opaque IndexError for an out-of-range topic.
• _is_btm_classifier() returned None instead of False.
• get_closest_topics() validated method only after loading the reference model.

Performance

• Fixed Jaccard topic matching and KL/Jeffrey divergence calculations.
• Corrected topic-distance embeddings and two-topic model handling.
• Improved comparisons across models with different vocabularies or topic counts.
• Fixed prevalence-weighted term probabilities and relevance rankings.
• Vectorized saliency calculations and optimized top-document selection.
• Strengthened input validation and optional model adapter handling.
• Improved interactive report reliability and caching.
• Added extensive regression coverage: 102 tests passing with 92% coverage.
• Added Python 3.9–3.14 CI and built-package validation.

• Fixed Jensen-Shannon and symmetric Kullback-Leibler divergence calculation (was using incorrect formulas)
• Fixed variable name bug in report interface
• Fixed sklearn MDS FutureWarning
• Added comprehensive input validation for probability matrices
• Cached phi matrix calculations to reduce redundant calls
• Added error handling for edge cases (division by zero, log(0))
• Added bounds checking and empty data validation
• Improved numerical stability with named constants
• Enhanced error messages for better debugging

This release fixed issues in marginal distribution and saliency calculations.

This release fixes a number of bugs in report method and slightly improves type hinting (thanks to @Nfsaavedra and @franciee2): #11, #13

This release fixes relevance calculation (#10, thanks to @ed9w2in6) and improves entropy test.

This release fixes the lower bound error in TSNE perplexity setting (#6, #9, thanks to @caicai555, @Yif18). It also includes minor code and package improvements.

This release fixes an error in get_phi method, adds some basic tests (to be added more in future), and updates packaging configs.

This release introduces improvements to entropy calculation method.

This release fixes a bug in topic scatter plot (#6) and optimizes t-SNE parameters for calculating topics 2D coordinates for scatter plot. It also includes minor cosmetic changes and fixes.