Topica: fast, all-in-one topic modeling for Python

topica is a fast, memory-efficient, all-in-one topic-modeling library for Python, built for social scientists. It brings together more than fifty models usually split across JVM tools like MALLET and R packages like stm, including LDA, STM, CTM, keyATM, BERTopic, and neural, dynamic, short-text, and embedding-based models, all under one NumPy-native API. Every model is validated against its reference implementation and reproducible from a fixed seed; all share one set of diagnostics, labeling, validation, and covariate-effect tools, so you learn a single workflow and it applies across the roster. It installs as a single wheel that needs only NumPy and pandas: no JVM, no PyTorch.

pip install topica

Quick start

Point topica at a DataFrame and read the topics. This runs exactly as written, on a bundled example dataset, right after install:

import topica

df = topica.datasets.load_gadarian()          # bundled; loads offline
corpus = topica.from_dataframe(
    df, text_col="open.ended.response", stopwords=topica.data.ENGLISH_STOPWORDS
)

model = topica.LDA(num_topics=5, seed=13)
model.fit(corpus)                             # sensible defaults; no tuning required
print(topica.summary(model))                  # top words per topic

from_dataframe keeps your metadata aligned to the documents that survive pruning, so the same corpus feeds a structural topic model that relates topic prevalence to a covariate and reports the effect with uncertainty:

prevalence = corpus.metadata[["treatment"]]   # a numeric DataFrame goes straight in

stm = topica.STM(num_topics=5, seed=13)
stm.fit(corpus, prevalence, prevalence_names=["treatment"])

draws  = topica.effects.posterior_theta_samples(stm, nsims=30, seed=0)
effect = topica.effects.estimate_effect(draws, prevalence, feature_names=["treatment"])

Your own data is one line away: pass pandas.read_csv("yours.csv") to from_dataframe. See the getting-started guide and the worked examples for analyses end to end.

Fits are reproducible and validated: the variational models are bit-exact against their references, the samplers reproduce from a fixed seed and thread count, and every model is checked against its reference implementation (R stm, MALLET, keyATM, and more).

The core needs only NumPy and pandas. Optional extras add features without weighing it down: topica[viz] (matplotlib plots), topica[formula] (R-style formulas), topica[polars] (Polars frames), and topica[llm] (LLM labels and embeddings, OpenAI or local via ollama).

Models

Find your goal in the left column. Every model is validated against a reference implementation, so the choice is about fit to your research design, not quality. A specialized model is often the right first choice when your data calls for it.

Common openings

If your goal is…	Start with	Also consider	First calls	Note
Explore themes with no prior structure	`LDA`	`NMF`	`search_k()`, `topic_table()`	The default first pass. `NMF` is a fast, deterministic alternative.
Relate topics to metadata (author, date, party)	`STM`	`DMR`	`estimate_effect()`, `one_hot()`, `spline()`	`STM` gives covariate effects with uncertainty; `DMR` is a lighter Gibbs prior.
Measure concepts you can name in advance	`KeyATM`	`SeededLDA`	`KeyATM(keywords=…)`, `.keyword_rate`	Anchor named topics with a few seed words each.
Very short documents: tweets, headlines, survey answers	`GSDMM`	`PT`	`fit()`	One topic per document; standard LDA over-fragments short text.
Cluster by meaning using embeddings	`BERTopic`	`ETM`	`fit(docs, doc_embeddings=…)`	Clustering, not a posterior: topic-proportion uncertainty and effect estimation behave differently than the models above.

Specialized approaches. Start here when your design calls for one.

If your data or goal is…	Start with	Also consider	First calls	Note
Topics shift over time slices	`DTM`	`DETM`	`fit(docs, times=…)`	Prevalence and content evolve across periods; `DETM` adds embeddings.
Documents linked in a network (citations, replies)	`RTM`	—	`fit(docs, links=…)`	Models the text and the link graph jointly.
Documents in more than one language	`PolylingualLDA`	—	`fit(doc_tuples)`	Aligned topics across languages from translation-linked tuples.
Place authors or actors on an ideological scale	`Wordfish`	`TBIP`	`fit(docs)`	Scaling from word usage; `TBIP` adds a text-based ideal-point prior.
How tone or sentiment varies with metadata	`STS`	—	`estimate_effect()`	Sentiment-discourse decomposition; reach for it when tone is the question.

topica.list_models(common_start=True) returns the common openings in code; list_models(group=…, brings=…) filters the full roster below.

All models (more than fifty, grouped by what you bring and what you want; click to expand)

Models are organized by what you bring and what you want, not by inference family. The from topica import X namespace is flat; topica.list_models(group=…, brings=…, inference=…, determinism=…) filters this roster in code. Brings is what you supply beyond raw text; Reproducibility is bit-exact (identical regardless of thread count), seed-reproducible (identical from a fixed seed and thread count), or llm-bounded.

Every model below is validated before it enters the roster: against a maintained reference implementation where one exists (MALLET, gensim, R stm, tomotopy, and the like), otherwise by planted recovery on a synthetic corpus with a known answer. See validation for where each model stands. The groupings are about fit to your research design, not quality: a specialized model is the right first choice when your data calls for it.

Common starting points

One per common goal, where most social scientists begin. BERTopic works differently from the others: it clusters document embeddings rather than fitting a posterior, so topic-proportion uncertainty and covariate-effect estimation do not carry over directly.

Model	Brings	Inference	Reproducibility	Summary
`LDA`	text	gibbs	seed-reproducible	Classic latent Dirichlet allocation via a fast SparseLDA collapsed-Gibbs sampler.
`NMF`	text	matrix-factorization	bit-exact	Non-negative matrix factorization of the document-term matrix via multiplicative updates.
`STM`	text, metadata	variational	bit-exact	Structural topic model: relate topic prevalence and content to covariates.
`KeyATM`	text, seeds	gibbs	seed-reproducible	Keyword-assisted topics: anchor named topics with a few seed words each.
`GSDMM`	text	gibbs	seed-reproducible	Gibbs-sampling Dirichlet mixture: one topic per short document.
`BERTopic`	text, embeddings	clustering	seed-reproducible	Cluster document embeddings; label topics by class-based TF-IDF.

Specialized approaches

The right first choice when your design calls for one: short text, change over time, document networks, multiple languages, ideological scaling, and more.

General-purpose

Model	Brings	Inference	Reproducibility	Summary
`OnlineLDA`	text	variational	seed-reproducible	Online (streaming) variational-Bayes LDA (Hoffman et al. 2010): minibatch stochastic VB with a decaying learning rate and a streaming partial_fit; the gensim LdaModel analogue for very large or streaming corpora.
`CTM`	text	variational	bit-exact	Correlated topic model: a logistic-normal prior that lets topics co-occur.
`ProdLDA`	text	vae	seed-reproducible	Product-of-experts LDA (AVITM) for sharper, more coherent topics; hand-coded VAE.
`HDP`	text	gibbs	seed-reproducible	Hierarchical Dirichlet process: infers the number of topics from the data.
`LSA`	text	svd	seed-reproducible	Latent semantic analysis: a truncated SVD of the weighted document-term matrix.
`AnchorLDA`	text	matrix-factorization	bit-exact	Anchor-words spectral recovery (Arora et al. 2013): deterministic, Gibbs-free topics from the word co-occurrence matrix.
`PolylingualLDA`	text	gibbs	seed-reproducible	Polylingual topic model (Mimno et al. 2009): aligned topics across languages from document tuples that share one topic distribution.
`CorEx`	text	information-theoretic	seed-reproducible	Correlation Explanation: information-theoretic topic model that maximizes total correlation; supports anchor words.
`MGLDA`	text	gibbs	seed-reproducible	Multi-Grain LDA: global (document-level) + local (sliding-window aspect) topics with a per-token grain switch. For reviews / aspect extraction.
`TopicalNGrams`	text	gibbs	seed-reproducible	Topical N-Grams (Wang, McCallum & Wei 2007): an LDA extension that jointly discovers topics and topic-specific multiword phrases. A per-token bigram-status indicator, sampled with the topic, decides whether a token continues a phrase from the previous word given its topic, so phrase structure is learned during fitting rather than fixed beforehand. Exposes top_phrases alongside top_words.

Covariates & structure

Model	Brings	Inference	Reproducibility	Summary
`STS`	text, metadata	variational	bit-exact	Structural topic-and-sentiment model over document metadata.
`SAGE`	text, metadata	gibbs	seed-reproducible	Sparse additive generative model: the same topic worded differently across groups.
`DMR`	text, metadata	gibbs	seed-reproducible	Dirichlet-multinomial regression: a document-metadata prior on topic proportions.
`GDMR`	text, metadata	gibbs	seed-reproducible	Generalized DMR with a smooth (Legendre-basis) prior over continuous covariates.
`Scholar`	text, metadata, labels	vae	seed-reproducible	SCHOLAR (Card et al. 2018): a ProdLDA VAE with a covariate-shifted prevalence prior, an optional supervised label head, and optional content (topic-covariate) word deviations — neural STM prevalence + sLDA + SAGE.
`RTM`	text, links	variational	seed-reproducible	Relational topic model (Chang & Blei 2010): jointly models document text and a link graph (citations, hyperlinks, adjacency); predicts links from words and words from links.
`FactorialLDA`	text	gibbs	seed-reproducible	Factorial LDA (Paul & Dredze 2012): each token is a K-tuple of latent factors (e.g. topic x sentiment); structured word priors tie tuples sharing a component and a sparsity prior deactivates unsupported tuples.
`AuthorTopic`	text, metadata	gibbs	seed-reproducible	Author-Topic Model: each author has a topic distribution; documents mix their authors. Answers what an author writes about.
`AuthorRecipientTopic`	text, metadata	gibbs	seed-reproducible	Author-Recipient-Topic (McCallum et al. 2007): topics conditioned on the (sender, recipient) pair, for the language of a directed social network (who talks to whom about what). Realized over the AuthorTopic engine.

Guided & supervised

Model	Brings	Inference	Reproducibility	Summary
`SeededLDA`	text, seeds	gibbs	seed-reproducible	Seeded LDA: steer named topics toward supplied seed words.
`GuidedNMF`	text, seeds	matrix-factorization	seed-reproducible	Guided NMF: seed-word-guided semi-supervised NMF; the matrix-factorization analogue of SeededLDA.
`LabeledLDA`	text, labels	gibbs	seed-reproducible	Labeled LDA: each document label is a topic; tokens are restricted to its labels.
`SupervisedLDA`	text, labels	variational	seed-reproducible	Supervised LDA: topics shaped to predict a per-document real-valued response.
`DiscLDA`	text, labels	gibbs	seed-reproducible	Discriminative LDA (Lacoste-Julien et al. 2008): topics split into per-class and shared blocks; reads how classes talk differently.

Short text

Model	Brings	Inference	Reproducibility	Summary
`PT`	text	gibbs	seed-reproducible	Pseudo-document topic model: pool short texts into pseudo-documents.
`BTM`	text	gibbs	seed-reproducible	Biterm topic model: learns topics from corpus-level word co-occurrence (biterms).

Dynamic & hierarchical

Model	Brings	Inference	Reproducibility	Summary
`DTM`	text, times	variational	seed-reproducible	Dynamic topic model: a fixed topic set whose word distributions drift across time slices.
`DETM`	text, embeddings, times	vae	seed-reproducible	Dynamic embedded topic model: embedding-factored topics that drift across time slices, fit as an amortized VAE.
`TopicsOverTime`	text, times	gibbs	seed-reproducible	Topics over Time: LDA with a per-topic Beta density over continuous timestamps; each topic has a temporal peak. Descriptive continuous-time prevalence (not vocabulary drift).
`HLDA`	text	gibbs	seed-reproducible	Hierarchical LDA (nested CRP): a learned tree of super- and sub-topics.
`PA`	text	gibbs	seed-reproducible	Pachinko allocation: a DAG of super- and sub-topics.

Embedding-based

Model	Brings	Inference	Reproducibility	Summary
`KeyNMF`	text, embeddings	matrix-factorization	bit-exact	KeyNMF (Kristensen-McLachlan et al. 2024): NMF over an embedding-derived keyword-importance matrix. For each document it scores its words by the similarity between the document embedding and the word embedding, keeps the top-N positive, and factors that sparse doc-word matrix. The bridge between the count-based NMF family and the embedding backend; sparse, readable topics robust to short/noisy text.
`Top2Vec`	text, embeddings	clustering	seed-reproducible	Topics as dense regions in a joint document-word embedding space.
`SemanticSignalSeparation`	text, embeddings	ica	seed-reproducible	Topics as independent axes of semantic space (S3, Kardos et al. 2025): FastICA over the document embeddings, with each word's importance read off by projecting the vocabulary embeddings onto each axis. Signed poles.
`ETM`	text, embeddings	variational	seed-reproducible	Embedded topic model: topic-word distributions factored through word embeddings.
`GaussianLDA`	text, embeddings	gibbs	seed-reproducible	Gaussian LDA (Das, Zaheer & Dyer 2015): each topic is a Gaussian over the word-embedding space (Normal-Inverse-Wishart prior), so topics generalize over semantically similar words. Collapsed Gibbs with a Student-t posterior predictive and rank-1 Cholesky up/downdates.
`FASTopic`	text, embeddings	optimal-transport	seed-reproducible	Topics from optimal-transport plans between document, topic, and word embeddings.
`CombinedTM`	text, embeddings	vae	seed-reproducible	Contextualized ProdLDA: encoder reads the bag of words plus a document embedding.
`ZeroShotTM`	text, embeddings	vae	seed-reproducible	Contextualized ProdLDA: encoder reads the document embedding alone, enabling cross-lingual transfer.
`InfoCTM`	text, dictionary	vae	seed-reproducible	Cross-lingual: two ProdLDA models aligned by a bilingual dictionary through a mutual-information term.

Ideal point

Model	Brings	Inference	Reproducibility	Summary
`Wordfish`	text	em	bit-exact	Poisson scaling (Slapin & Proksch 2008): an unsupervised one-dimensional ideal-point estimate from word frequencies alone, no topics. The word-frequency baseline companion to IdealPointTM.
`Wordshoal`	text, metadata	em	bit-exact	Multi-domain scaling (Lauderdale & Herzog 2016): scales each debate/domain with Wordfish, then combines the within-domain positions into one cross-domain actor scale via a linear factor model. The multi-domain extension of Wordfish, for speeches carrying trusted debate labels.
`TBIP`	text	variational	seed-reproducible	Text-Based Ideal Points (Vafa, Naidu & Blei 2020): a Poisson factorization whose neutral topic-word intensities are rescaled by a per-word ideological factor exp(x_s * eta_kv), with the author position x_s latent. Fit by the paper's mean-field variational inference (reparameterized SVI). Recovers ideological scales from unlabeled text.
`PartyEmbeddings`	text, metadata	neural-embedding	seed-reproducible	Party embeddings (Rheault & Cochrane 2020): a PV-DM paragraph-vector model trained by negative sampling with party-period metadata tags; the leading principal components of the learned party vectors give the ideological scale, and words share the space so a party's language can be read off by proximity. The corpus-trained word-embedding member of the ideal-point family.

LLM-based

Model	Brings	Inference	Reproducibility	Summary
`TopicGPT`	text, llm	prompting	llm-bounded	LLM-driven topic discovery: prompt a model to propose, refine, and assign a topic taxonomy with descriptions.

Experimental

Not (yet) on the validated roster, on one of two grounds: the model is unpublished (a topica original with no paper), or it is a published method whose benefit has not held up against a simpler baseline. A published method with faithful inference stays validated even when its basis is planted-recovery, so planted validation alone does not land a model here. Gated: call topica.enable_experimental() (or set TOPICA_EXPERIMENTAL=1) before use. These may change or be removed without a deprecation cycle. For the triple gate a model clears to graduate to the validated roster, and where every model stands on validation evidence, see Validation & graduation.

Model	Brings	Inference	Reproducibility	Summary
`TensorLDA`	text	svd	seed-reproducible	Online Tensor LDA (Kangaslahti et al. 2026): deterministic method-of-moments topic modeling via second and third-order cumulants.
`NarrativeTM`	text	gibbs	seed-reproducible	Intra-document narrative trajectory model: captures how topic prevalence shifts across the progress of a text.
`CSATM`	text, links	gibbs	seed-reproducible	Conversational Structure Aware TM (Sun et al. 2020): weights each comment's tokens by a reply-tree 'popularity' score and, after Gibbs, smooths each comment's topics toward its ancestors along the reply path ('transitivity'). For threaded forum data (posts + nested comments). Ported from the paper (no reference implementation); validated by planted recovery + LDA reduction.
`ThreadTM`	text, links	variational	seed-reproducible	ThreadTM (topica-original): a reply-threaded topic model — CTM logistic-normal topics with a reply-tree structured prior, so a reply's topic prior is coupled to the comment it answers (a persistence-smoothing prior, reverting toward its covariate-group baseline). Supports a prevalence covariate and an optional content covariate (topic words shift by depth/level, à la STM's content model). For threaded discussion (posts + nested comments). Validated by planted recovery and a synthetic held-out-beat gate (the parent's topics predict held-out leaf tokens better than the no-tree baseline on persistence-structured data). The real-corpus benefit is genre-dependent; read kappa_ci before claiming persistence.
`IdealPointTM`	text, embeddings	variational	seed-reproducible	Topic model with a latent ideal-point head: each author gets a low-dimensional position that shifts within-topic word choice, with a per-topic discrimination. Consumes word tokens as counts (Wordfish with topics) or, when word embeddings are supplied to fit, factored through them as in ETM. The unsupervised, latent-trait twin of the STM content covariate.
`IdealPointSentenceTM`	text, embeddings	em	seed-reproducible	Continuous ideal-point topic model over sentence/document embeddings: topics are Gaussian clusters whose centroids are displaced by a latent author position. The sentence-embedding sibling of IdealPointTM, fit by EM.
`EmbeddingLDA`	text, embeddings	gibbs	seed-reproducible	LDA anchored by pre-trained embeddings: k-means clusters the vocabulary embeddings, seeds each topic with the words nearest a cluster centroid, and (optionally) biases each document's mixture toward its own embedding. A topica original; validated by planted-recovery only.

Every model exposes the same shape: fit(docs, …), then topic_word (φ), doc_topic (θ), top_words(n), and save/load, so one diagnostic, labeling, and effect-estimation stack applies to all of them and a new model inherits it for free. The embedding-based models take document vectors from any embedder (sentence-transformers, an API, or a local model such as ollama; no PyTorch or UMAP/numba in the wheel). Full guides: the models and embedding topics.

Diagnostics & analysis

Model-agnostic: they work on any fitted model's topic_word/doc_topic:

Quality: coherence (u_mass, c_v, c_uci, c_npmi; co-occurrence counting in the Rust core), exclusivity, topic_diversity, quality_frontier
Labeling: label_topics (prob / FREX / lift / score), frex, relevance, find_thoughts, topic_table, summary
Validation: word_intrusion, document_intrusion, bootstrap_stability, search_k
Reliability: select_model (fit many seeds) and ensemble (combine runs into a consensus more reliable than any single fit — cluster/align/stable methods, the last derived from gensim's EnsembleLda)
Comparison: fighting_words (weighted log-odds) for contrasting corpora
Covariate effects: estimate_effect (method of composition, cluster-robust SEs, GLM links), topic_correlation, and the design helpers one_hot, spline, and interaction (in topica.effects / topica.design; they build covariate bases for any model's design matrix); posterior_theta_samples draws θ for the logistic-normal models (STM/CTM)
Preprocessing: tokenize, learn_phrases / apply_phrases, split_documents, the Corpus class

See diagnostics and covariate effects.

Performance

topica runs on a parallel Rust core, so the whole roster is fast and every fit is reproducible from a fixed seed. Core for core it matches the hand-tuned compiled samplers: parity with Java MALLET on plain LDA and with the C++ keyATM on keyword models. For the structural and other variational models it is several times faster than R stm, the single-threaded field standard. Fit to convergence (both at the same emtol, spectral start), on real corpora:

Model	Reference	topica speedup (to convergence)
STM	R `stm`	1.7–2.7× single-threaded, ~5–7× multicore
LDA	Java MALLET	parity single-threaded; multithread speedup grows with corpus size
keyATM	R `keyATM`	parity single-threaded, ~2× multithreaded

topica also fits in about a quarter of R stm's memory (≈180MB against ≈675MB at 5,000 documents). For the approximate parallel Gibbs samplers the multithreaded speedup grows with corpus size: the per-sweep count-table merge is fixed overhead, so larger corpora amortize it over more sampling work. LDA's eight-core speedup over MALLET runs about 3× at 2,000 documents and reaches ~4× at 5,000.

Every fit is reproducible from a fixed seed and validated against its reference. See Benchmarks for the full methodology; reproduce the structural-model table with python benchmarks/bench_stm_convergence.py and the size-varying LDA curve with python benchmarks/speed_vs_size.py.

Install from source

pip install maturin
git clone https://github.com/nealcaren/topica && cd topica
python -m venv .venv && source .venv/bin/activate
maturin develop --release --features python

Requires numpy >= 1.21. Use --release (the debug build is much slower).

Acknowledgements

Topica was inspired by a post from David Mimno about porting Java MALLET to Rust. As a long-time Python user, I had long been jealous of the topic-modeling tools available in other languages; this seemed like an opportunity to make those capabilities easier to use in Python for me and for others.

As such, Topica stands on a generation of open topic-modeling research and code. Each entry below lists the reference, its authors and year, and the topica class(es) it underlies; the other models are Rust ports or reimplementations, validated against these reference implementations.

MALLET (McCallum, 2002) — LDA, DMR, LabeledLDA: the SparseLDA sampler, Dirichlet-multinomial regression, and hyperparameter optimization. LDA began as a port of David Mimno's RustMallet (Apache-2.0) and follows its SparseLDA sampler and fixed-point optimizer closely, but uses its own RNG (PCG), so it is not byte-identical to RustMallet. Against Java MALLET (also a different RNG) it recovers the same topics on a planted corpus (cosine 1.000)
stm (Roberts, Stewart & Tingley, 2019) — STM, CTM, SAGE: variational EM, estimateEffect, searchK, FREX, spectral initialization, and the method of composition
sts (Chen & Mankad, 2024) — STS: the Structural Topic and Sentiment-Discourse model — the joint prevalence/sentiment Laplace E-step and the Poisson topic-word M-step, validated against the package
lda-c / ctm-c / dtm and hdp (Blei lab, 2006–2007) — CTM, DTM, HDP: the CTM, Dynamic Topic Model, and HDP samplers
gensim (Řehůřek & Sojka, 2010) — DTM, ensemble, OnlineLDA: the coherence-pipeline conventions (the coherence_type= API and default sliding windows; the measures themselves are Röder et al. 2015 and Mimno et al. 2011), the LdaSeqModel DTM reference, the EnsembleLda (CBDBSCAN stable-topic) method that ensemble(method="stable") derives from (matching it on well-separated inputs, with two documented improvements on edge cases where gensim degenerates), and the LdaModel online-VB reference for OnlineLDA
onlineldavb (Hoffman, Blei & Bach, 2010) — OnlineLDA: online (streaming) variational Bayes for LDA — the minibatch stochastic-VB E-step, the decaying Robbins-Monro learning rate, and the streaming partial_fit; written from the paper and validated against onlineldavb.py and gensim's LdaModel as external oracles (both are copyleft, so no code was copied)
tomotopy (bab2min, 2020) — API conventions (summary, the short-text models), and GDMR (generalized DMR; Lee & Song, 2020), validated against its GDMRModel
scikit-learn (Pedregosa et al., 2011) — NMF: the multiplicative-update solver (Lee & Seung, 2001) and the NNDSVD initialization (Boutsidis & Gallopoulos, 2008), validated against sklearn.decomposition.NMF (BSD-3-Clause); and LSA: latent semantic analysis / indexing (Deerwester et al., 1990), validated against sklearn.decomposition.TruncatedSVD (BSD-3-Clause) including its svd_flip sign convention. The numerics are reimplemented in Rust; the randomized truncated SVD shared by both (it seeds NMF's NNDSVD and is the LSA factorization itself) follows Halko et al. (2011).
keyATM (Eshima, Imai & Sasaki, 2024) — KeyATM: the base, covariate, and dynamic models, the information-theory token weighting, and the Chib (1998) change-point HMM, validated against the package
seededlda (Watanabe, 2023) — SeededLDA: the corpus-frequency-scaled seed prior (count × weight × 100), validated against the package's seed matrix and seeded topics
LightLDA (Yuan et al., 2015) — LDA: the alias-table Metropolis-Hastings sampler
GSDMM (Yin & Wang, 2014) — GSDMM: the movie-group-process mixture for short text
BTM (Yan, Guo, Lan & Cheng, 2013; R package by Jan Wijffels) — BTM: the biterm co-occurrence topic model for short text
Polylingual Topic Models (Mimno, Wallach, Naradowsky, Smith & McCallum, 2009) — PolylingualLDA: LDA over aligned document tuples that share one topic distribution, giving topics aligned across many languages; validated against MALLET's PolylingualTopicModel as a black-box oracle
DiscLDA (Lacoste-Julien, Sha & Jordan, 2008) — DiscLDA: discriminative LDA with per-class and shared topic blocks; the fixed block-transform variant, validated against the paper's 20 Newsgroups feature-classification result (no reference implementation exists, so it is paper-derived)
Factorial LDA (Paul & Dredze, 2012) — FactorialLDA: sparse multi-dimensional topics, where each token is a K-tuple of latent factors tied by structured log-linear priors; implemented from the paper's mathematics (the reference Java is GPL and non-reproducible, so the port is certified by finite-difference gradient and factor-tying tests plus planted recovery, not seed parity)
ProdLDA / AVITM (Srivastava & Sutton, 2017) — ProdLDA: autoencoding variational inference and the product-of-experts word model
SCHOLAR (Card, Tan & Smith, 2018; reference dallascard/scholar, Apache-2.0) — Scholar: metadata in a ProdLDA VAE — a covariate-dependent topic-prevalence prior (neural STM prevalence), an optional supervised label head (neural sLDA), and optional content/topic-covariate word deviations (neural SAGE), on topica's ProdLDA backbone, validated against the reference as a numerical oracle
BERTopic (Grootendorst, 2022) and Top2Vec (Angelov, 2020) — BERTopic, Top2Vec: the embedding-clustering pipeline, class-based TF-IDF, and the reduce → cluster → represent design
CETopic / topicx (Zhang, Fang, Chen & Namazi-Rad, NAACL 2022, MIT) — BERTopic(weighting="tfidf-idf"): the TFIDF×IDF_i topic-word selection scheme (a corpus-level TF-IDF averaged per cluster times a cross-cluster IDF penalty), ported faithfully to the reference's scikit-learn defaults
S³ / turftopic (Kardos, Kostkan, Enevoldsen, Vermillet, Nielbo & Rocca, ACL 2025, MIT) — SemanticSignalSeparation: Semantic Signal Separation, FastICA over contextual document embeddings with topic words read off by projecting the vocabulary embeddings onto each independent axis, ported faithfully to the reference's scikit-learn FastICA defaults
ETM (Dieng, Ruiz & Blei, 2020) — ETM: the Embedded Topic Model (per-document variational EM and an amortized VAE)
DETM (Dieng, Ruiz & Blei, 2019) — DETM: the Dynamic Embedded Topic Model (structured amortized variational inference with a hand-coded LSTM)
FASTopic (Wu et al., 2024) — FASTopic: the optimal-transport topic model
contextualized-topic-models (Bianchi et al., MIT) — CombinedTM (Bianchi, Terragni & Hovy, 2021) and ZeroShotTM (Bianchi, Nozza & Hovy, 2021): ProdLDA encoders that read a contextual document embedding, alongside or in place of the bag of words
quanteda.textmodels (Benoit et al., 2018) — Wordfish: the Slapin & Proksch (2008) Poisson scaling model, validated against its textmodel_wordfish (the recovered scale and the analytic position standard errors both match at correlation 1.00 on a corpus sampled from the model)
tbip (Vafa, Naidu & Blei, 2020) — TBIP: Text-Based Ideal Points; the official implementation is TensorFlow 1.x, so topica reimplements the published model and its mean-field variational inference, validated against an independent PyTorch reference
TensorLy TLDA (Kangaslahti, Ebanks, Kossaifi, Liu, Alvarez & Anandkumar, 2026) — TensorLDA: online tensor latent Dirichlet allocation via second- and third-order moments. The Rust implementation is experimental; see the TensorLDA validation record for its current evidence and limitations.
partyembed (Rheault & Cochrane, 2020) — PartyEmbeddings: party embeddings via a PV-DM paragraph-vector model with party-period metadata tags, placed by PCA of the learned party vectors. The reference builds on gensim's Doc2Vec; topica reimplements the PV-DM negative-sampling training in Rust (from Mikolov et al. 2013 and Le & Mikolov 2014) and is validated against that Doc2Vec scale (correlation 1.00 on a planted ordering)
CLNTM (Nguyen & Luu, 2021) — the InfoNCE contrastive regularization on topic vectors offered by the contrastive= flag on the VAE models
WHAI / Weibull-Dirichlet VAE (Zhang et al., 2018; Burkhardt & Kramer, 2019) — the Weibull-reparameterized Dirichlet prior offered by prior="dirichlet" on the VAE models
Neural variational topic models with alternative priors (Miao, Grefenstette & Blunsom, 2017; Nalisnick & Smyth, 2017) — the Gaussian stick-breaking prior offered by prior="stick_breaking" on the VAE models
TopicGPT (Pham et al., NAACL 2024, MIT) — TopicGPT: the generate / refine / assign prompt flow for LLM-driven topic discovery

The embedding-native models build on two pure-Rust crates: petal-clustering for HDBSCAN and umap-rs for the optional UMAP reducer, both BLAS-free.

Full citations for every model and reference implementation, and how to cite topica, are on the Citing page.

Contributing, tests, and support

Contributions are welcome. See CONTRIBUTING for the development setup and workflow, CONTRIBUTING-MODELS for adding a new topic model, and the conventions guide for the cross-model naming and API contract. All participants are expected to follow our Code of Conduct.

To run the test suite after a source build:

cargo test --lib                                   # Rust unit tests
python -m pytest tests/ -q                         # Python tests
mkdocs build --strict                              # docs build clean

Every push runs these on CI (see the badge above). The parity/ checks validate models against their reference implementations (R stm, keyATM, MALLET); they skip cleanly when those toolchains are not installed.

Report a bug or request a feature: open an issue.
Ask a question or share how you are using topica: start a discussion.

topica is maintained by Neal Caren. Issues and pull requests are triaged on a best-effort basis.

Citation

If you use topica in published work, please cite it. GitHub's Cite this repository button (top right) generates a formatted reference from CITATION.cff. A software paper is in preparation; until it appears, cite the software release:

@software{caren_topica,
  author  = {Caren, Neal},
  title   = {topica: fast, all-purpose topic modeling for Python},
  year    = {2026},
  url     = {https://github.com/nealcaren/topica},
  version = {0.54.0}
}

Replace version with the release you used. For the individual models and their reference implementations, see the Citing page.

License

Apache-2.0 — see LICENSE.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

topica-0.58.0.tar.gz (20.3 MB view details)

Uploaded Sep 3, 2026 Source

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

topica-0.58.0-cp39-abi3-win_amd64.whl (5.4 MB view details)

Uploaded Sep 3, 2026 CPython 3.9+Windows x86-64

topica-0.58.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (5.2 MB view details)

Uploaded Sep 3, 2026 CPython 3.9+manylinux: glibc 2.17+ x86-64

topica-0.58.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (4.9 MB view details)

Uploaded Sep 3, 2026 CPython 3.9+manylinux: glibc 2.17+ ARM64

topica-0.58.0-cp39-abi3-macosx_11_0_arm64.whl (4.7 MB view details)

Uploaded Sep 3, 2026 CPython 3.9+macOS 11.0+ ARM64

topica-0.58.0-cp39-abi3-macosx_10_12_x86_64.whl (5.0 MB view details)

Uploaded Sep 3, 2026 CPython 3.9+macOS 10.12+ x86-64

File details

Details for the file topica-0.58.0.tar.gz.

File metadata

Download URL: topica-0.58.0.tar.gz
Upload date: Sep 3, 2026
Size: 20.3 MB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for topica-0.58.0.tar.gz
Algorithm	Hash digest
SHA256	`73063a797cf81006d37d03448f45c85ea524b8a4d2bd681bfe6bd7a77d2ce9f7`
MD5	`370f48818f3eca59fa0c155eefd892ef`
BLAKE2b-256	`25036e67e5bbeba4deb9bcdf2f66e0f2feea5c2ae829c9676336d1ec1ca3823a`

See more details on using hashes here.

Provenance

The following attestation bundles were made for topica-0.58.0.tar.gz:

Publisher: CI.yml on nealcaren/topica

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: topica-0.58.0.tar.gz
- Subject digest: 73063a797cf81006d37d03448f45c85ea524b8a4d2bd681bfe6bd7a77d2ce9f7
- Sigstore transparency entry: 2704792945
- Sigstore integration time: Sep 3, 2026
Source repository:
- Permalink: nealcaren/topica@fc22b0b64db353163bb201cc182ea04bae7272d7
- Branch / Tag: refs/tags/v0.58.0
- Owner: https://github.com/nealcaren
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: CI.yml@fc22b0b64db353163bb201cc182ea04bae7272d7
- Trigger Event: push

File details

Details for the file topica-0.58.0-cp39-abi3-win_amd64.whl.

File metadata

Download URL: topica-0.58.0-cp39-abi3-win_amd64.whl
Upload date: Sep 3, 2026
Size: 5.4 MB
Tags: CPython 3.9+, Windows x86-64
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for topica-0.58.0-cp39-abi3-win_amd64.whl
Algorithm	Hash digest
SHA256	`aefbba1af96bf3e4ea58339cd3195b98f768d943e9acc80aedee55282b0bc890`
MD5	`6935b9ae8892bf99507578e575e00da6`
BLAKE2b-256	`6a1e64b5af40bd6db73780ae5d6e8efbe02735e243dc9843192b090366b462e4`

See more details on using hashes here.

Provenance

The following attestation bundles were made for topica-0.58.0-cp39-abi3-win_amd64.whl:

Publisher: CI.yml on nealcaren/topica

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: topica-0.58.0-cp39-abi3-win_amd64.whl
- Subject digest: aefbba1af96bf3e4ea58339cd3195b98f768d943e9acc80aedee55282b0bc890
- Sigstore transparency entry: 2704793138
- Sigstore integration time: Sep 3, 2026
Source repository:
- Permalink: nealcaren/topica@fc22b0b64db353163bb201cc182ea04bae7272d7
- Branch / Tag: refs/tags/v0.58.0
- Owner: https://github.com/nealcaren
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: CI.yml@fc22b0b64db353163bb201cc182ea04bae7272d7
- Trigger Event: push

File details

Details for the file topica-0.58.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.

File metadata

Download URL: topica-0.58.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Upload date: Sep 3, 2026
Size: 5.2 MB
Tags: CPython 3.9+, manylinux: glibc 2.17+ x86-64
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for topica-0.58.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Algorithm	Hash digest
SHA256	`3ab32e1c3467d807ed37aaad30f4f601b3474468dc078a31e22e61fb4874aed1`
MD5	`3a166b5a3c02f84aa0e2647ee1bc5a85`
BLAKE2b-256	`cb6ad3ffe97288797f6fbb0775dd93f14d1d103396edfb8930b5c3091920f6db`

See more details on using hashes here.

Provenance

The following attestation bundles were made for topica-0.58.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl:

Publisher: CI.yml on nealcaren/topica

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: topica-0.58.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
- Subject digest: 3ab32e1c3467d807ed37aaad30f4f601b3474468dc078a31e22e61fb4874aed1
- Sigstore transparency entry: 2704793173
- Sigstore integration time: Sep 3, 2026
Source repository:
- Permalink: nealcaren/topica@fc22b0b64db353163bb201cc182ea04bae7272d7
- Branch / Tag: refs/tags/v0.58.0
- Owner: https://github.com/nealcaren
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: CI.yml@fc22b0b64db353163bb201cc182ea04bae7272d7
- Trigger Event: push

File details

Details for the file topica-0.58.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl.

File metadata

Download URL: topica-0.58.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
Upload date: Sep 3, 2026
Size: 4.9 MB
Tags: CPython 3.9+, manylinux: glibc 2.17+ ARM64
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for topica-0.58.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
Algorithm	Hash digest
SHA256	`d69f66bf9e228b3f8e31ca86f0ae42a98c75539e64d066269a92e632ee427d12`
MD5	`bd00c0cb27f761a4fe066d1719d65131`
BLAKE2b-256	`e673d91024def77bc4f3e60be57b32ce89f6c816d461a1b70f5cfb3d9f342756`

See more details on using hashes here.

Provenance

The following attestation bundles were made for topica-0.58.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl:

Publisher: CI.yml on nealcaren/topica

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: topica-0.58.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
- Subject digest: d69f66bf9e228b3f8e31ca86f0ae42a98c75539e64d066269a92e632ee427d12
- Sigstore transparency entry: 2704793045
- Sigstore integration time: Sep 3, 2026
Source repository:
- Permalink: nealcaren/topica@fc22b0b64db353163bb201cc182ea04bae7272d7
- Branch / Tag: refs/tags/v0.58.0
- Owner: https://github.com/nealcaren
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: CI.yml@fc22b0b64db353163bb201cc182ea04bae7272d7
- Trigger Event: push

File details

Details for the file topica-0.58.0-cp39-abi3-macosx_11_0_arm64.whl.

File metadata

Download URL: topica-0.58.0-cp39-abi3-macosx_11_0_arm64.whl
Upload date: Sep 3, 2026
Size: 4.7 MB
Tags: CPython 3.9+, macOS 11.0+ ARM64
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for topica-0.58.0-cp39-abi3-macosx_11_0_arm64.whl
Algorithm	Hash digest
SHA256	`f9037660fa1a7dc43ca48ea8867fec1e23c5684c40db9ea40aeccd462b3f5936`
MD5	`1401136529a278f1887123d401867293`
BLAKE2b-256	`059b39e873ad609e00a7cc1d2ca8b5a0bfaff837fcb35e1a0bf6d9b920dac849`

See more details on using hashes here.

Provenance

The following attestation bundles were made for topica-0.58.0-cp39-abi3-macosx_11_0_arm64.whl:

Publisher: CI.yml on nealcaren/topica

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: topica-0.58.0-cp39-abi3-macosx_11_0_arm64.whl
- Subject digest: f9037660fa1a7dc43ca48ea8867fec1e23c5684c40db9ea40aeccd462b3f5936
- Sigstore transparency entry: 2704793003
- Sigstore integration time: Sep 3, 2026
Source repository:
- Permalink: nealcaren/topica@fc22b0b64db353163bb201cc182ea04bae7272d7
- Branch / Tag: refs/tags/v0.58.0
- Owner: https://github.com/nealcaren
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: CI.yml@fc22b0b64db353163bb201cc182ea04bae7272d7
- Trigger Event: push

File details

Details for the file topica-0.58.0-cp39-abi3-macosx_10_12_x86_64.whl.

File metadata

Download URL: topica-0.58.0-cp39-abi3-macosx_10_12_x86_64.whl
Upload date: Sep 3, 2026
Size: 5.0 MB
Tags: CPython 3.9+, macOS 10.12+ x86-64
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for topica-0.58.0-cp39-abi3-macosx_10_12_x86_64.whl
Algorithm	Hash digest
SHA256	`4058d4e1dec52fd8533df9d7cd01c9b42ead19e4fa160d0d3f76d115827e9963`
MD5	`c21766f895762ab3221661a5151496e7`
BLAKE2b-256	`19da3983b7d5ed2e21f038058dbf389e975fba7d135194fc0de8c86904c27a2c`

See more details on using hashes here.

Provenance

The following attestation bundles were made for topica-0.58.0-cp39-abi3-macosx_10_12_x86_64.whl:

Publisher: CI.yml on nealcaren/topica

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: topica-0.58.0-cp39-abi3-macosx_10_12_x86_64.whl
- Subject digest: 4058d4e1dec52fd8533df9d7cd01c9b42ead19e4fa160d0d3f76d115827e9963
- Sigstore transparency entry: 2704793093
- Sigstore integration time: Sep 3, 2026
Source repository:
- Permalink: nealcaren/topica@fc22b0b64db353163bb201cc182ea04bae7272d7
- Branch / Tag: refs/tags/v0.58.0
- Owner: https://github.com/nealcaren
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: CI.yml@fc22b0b64db353163bb201cc182ea04bae7272d7
- Trigger Event: push