Agent cheat sheet

A one-screen reference for an LLM agent or a first-time user who has just run import topica and needs the canonical patterns without reading the whole documentation set. Everything below is also available at the REPL:

import topica

topica.guide()            # the one-screen essentials, printed
topica.guide("STM")       # one model: purpose, signatures, first calls
topica.guide(full=True)   # every validated model, grouped

topica.guide() renders live from the model registry and each model's real signature, so it always matches the installed build. It ships in the wheel, so it is available from a plain pip install topica with no repository checkout. The page below is generated from the same builder; do not edit it by hand. Edit python/topica/_guide.py and run scripts/gen_guide.py.

For task-shaped starting points you can transplant onto your own data (compare prevalence across groups, compare how groups word a topic, track topics over time), see the task recipes in examples/recipes/.

The one-screen guide

topica - quick guide for agents and first-time users

THE WORKFLOW (a full LDA analysis in five lines)
    import topica
    corpus = topica.from_dataframe(df, text_col="text")      # build + prune vocab
    res    = topica.select.search_k(corpus, ks=[10, 20, 30])  # choose K; res.best_k()
    model  = topica.LDA(num_topics=res.best_k()).fit(corpus)
    topica.inspect.topic_table(model)                         # labelled topics
    # with metadata:
    topica.effects.estimate_effect(model, X=X, corpus=corpus) # covariate effects + CIs

EVERY FITTED MODEL EXPOSES THE SAME SURFACE
    model.topic_word                           (K, V) topic-word matrix; rows sum to 1 (generative models)
    model.doc_topic                            (D, K) document-topic matrix; rows sum to 1
    model.vocabulary                           V words, aligned to topic_word columns
    model.top_words(n)                         list[list[str]] top n per topic; weights=True for (word, prob)
    model.num_topics / .doc_names / .settings  K, row labels, and the fit config as a dict
    model.save(path) / Model.load(path)        round-trip a fitted model to disk

PICK A MODEL BY GOAL
  Common openings:
    Explore themes with no prior structure
        -> LDA (or NMF): `search_k()`, `topic_table()`
    Relate topics to metadata (author, date, party)
        -> STM (or DMR): `estimate_effect()`, `one_hot()`, `spline()`
    Measure concepts you can name in advance
        -> KeyATM (or SeededLDA): `KeyATM(keywords=…)`, `.keyword_rate`
    Very short documents: tweets, headlines, survey answers
        -> GSDMM (or PT): `fit()`
    Cluster by meaning using embeddings
        -> BERTopic (or ETM): `fit(docs, doc_embeddings=…)`
  Specialized (start here when your design calls for it):
    Topics shift over time slices
        -> DTM (or DETM): `fit(docs, times=…)`
    Documents linked in a network (citations, replies)
        -> RTM: `fit(docs, links=…)`
    Documents in more than one language
        -> PolylingualLDA: `fit(doc_tuples)`
    Place authors or actors on an ideological scale
        -> Wordfish (or TBIP): `fit(docs)`
    How tone or sentiment varies with metadata
        -> STS: `estimate_effect()`

HELPER NAMESPACES (topica.<stage>.*)
    select                             choosing K (search_k, select_model)
    inspect                            reading topics (label_topics, topic_table, frex, find_thoughts)
    evaluate                           validation (coherence, exclusivity, topic_stability, perplexity)
    effects                            covariate effects (estimate_effect, predicted_prevalence)
    design                             design matrices (one_hot, design_matrix, spline)
    data                               corpus + bundled datasets (from_dataframe, tokenize, datasets)
    compare / provenance / embeddings  two-fit drift, analysis manifest, embedding I/O
    -> guide("<name>") prints any helper's signature (e.g. guide("estimate_effect"))

GO DEEPER
    topica.guide("STM")            one model: signatures + first calls
    topica.guide("topic_stability") one helper: signature + purpose
    topica.guide(full=True)        every model, grouped
    help(topica.STM)               full constructor / fit docstrings
    topica.list_models()           the roster (list_models(group=...) to filter)
    docs: https://nealcaren.github.io/topica/

Every validated model

topica model reference (validated roster)

### General-purpose

LDA(num_topics, *, alpha_sum=None, beta=0.01, optimize_interval=50, burn_in=200, seed=13, num_threads=1, sampler='sparse', mh_steps=2, use_symmetric_alpha=False, init='random')
    Classic latent Dirichlet allocation via a fast SparseLDA collapsed-Gibbs sampler.
    .fit(data, *, iters=1000, num_samples=5, sample_interval=25, progress=None, progress_interval=50, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10, num_threads=None, turbo_merge_every=1)

OnlineLDA(num_topics, *, alpha_sum=None, beta=0.01, tau=1.0, kappa=0.7, batch_size=256, inner_iters=100, mean_change_tol=0.001, total_docs=None, seed=13)
    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.
    .fit(data, *, iters=100, convergence_tol=0.0, progress=None)

CTM(num_topics, *, sigma_shrink=0.0, seed=13, init='spectral', variational='laplace')
    Correlated topic model: a logistic-normal prior that lets topics co-occur.
    .fit(data, *, iters=500, convergence_tol=1e-05, inference='batch', batch_size=256, tau=64.0, kappa=0.7, beta_init=None, em_tol=None, keep_eta_cov=True, num_threads=None, spectral_projection_threshold=Ellipsis, progress=None)

ProdLDA(num_topics, *, alpha=1.0, hidden_size=100, dropout=0.2, batch_size=200, lr=0.002, convergence_tol=0.0, seed=13, prior=Ellipsis, contrastive=False, contrastive_weight=0.5, contrastive_temp=0.5, em_tol=None)
    Product-of-experts LDA (AVITM) for sharper, more coherent topics; hand-coded VAE.
    .fit(data, *, iters=None, convergence_tol=None, progress=None)

HDP(*, alpha=0.1, gamma=0.1, beta=0.01, seed=13, resample_conc=True, concentration_max=1000000.0, eta=None)
    Hierarchical Dirichlet process: infers the number of topics from the data.
    .fit(data, *, iters=150, progress_interval=0, keep_theta_draws=True, num_theta_draws=25, report_interval=None)

NMF(num_topics, *, beta_loss='frobenius', init='nndsvd', weighting='tfidf', convergence_tol=0.0001, seed=13)
    Non-negative matrix factorization of the document-term matrix via multiplicative updates.
    .fit(data, *, iters=None, convergence_tol=None, num_threads=None)

LSA(num_topics, *, weighting='tfidf', seed=13)
    Latent semantic analysis: a truncated SVD of the weighted document-term matrix.
    .fit(data, *, num_threads=None)

AnchorLDA(num_topics: 'int', *, recover: 'str' = 'kl', min_count: 'int' = 5, seed: 'int' = 13, eta: 'float' = 1.0, convergence_tol: 'float' = 1e-05, frex_w: 'float' = 0.5, frequency_temper: 'float' = 0.5, anchor_min_doc_freq: 'float' = 0.01)
    Anchor-words spectral recovery (Arora et al. 2013): deterministic, Gibbs-free topics from the word co-occurrence matrix.
    .fit(data, *, iters=None, min_count=None)

PolylingualLDA(num_topics, *, alpha=None, beta=0.01, iters=1000, optimize_alpha=True, optimize_interval=10, optimize_burn_in=200, seed=13)
    Polylingual topic model (Mimno et al. 2009): aligned topics across languages from document tuples that share one topic distribution.
    .fit(data, *, iters=None, progress=None)

CorEx(num_topics=2, *, anchor_words=None, anchor_strength=1.0, count='binarize', convergence_tol=1e-05, seed_match='fixed', case_insensitive=False, seed=13)
    Correlation Explanation: information-theoretic topic model that maximizes total correlation; supports anchor words.
    .fit(data, *, iters=None, convergence_tol=None, num_threads=None)

MGLDA(num_global_topics, num_local_topics, *, window=3, alpha_global=0.1, alpha_local=0.1, alpha_mix_global=0.1, alpha_mix_local=0.1, beta_global=0.01, beta_local=0.01, gamma=0.1, seed=13)
    Multi-Grain LDA: global (document-level) + local (sliding-window aspect) topics with a per-token grain switch. For reviews / aspect extraction.
    .fit(data, *, iters=1000, progress=None)

TopicalNGrams(num_topics, *, alpha_sum=50.0, beta=0.01, gamma=0.01, delta1=1.0, delta2=1.0, min_count=1, seed=13)
    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.
    .fit(data, *, iters=1000, progress=None)

### Covariates & structure

STM(num_topics, *, sigma_shrink=0.0, seed=13, init='spectral', variational='laplace')
    Structural topic model: relate topic prevalence and content to covariates.
    .fit(data, prevalence=None, *, prevalence_names=None, content=None, content_names=None, content_time=None, content_smooth=1.0, content_prior_var=0.5, content_prior='l2', iters=500, convergence_tol=1e-05, gamma_prior='pooled', gamma_enet=1.0, beta_init=None, em_tol=None, covariates=None, keep_eta_cov=True, num_threads=None, spectral_projection_threshold=Ellipsis, progress=None)

STS(num_topics, *, seed=13, init='spectral')
    Structural topic-and-sentiment model over document metadata.
    .fit(data, sentiment_seed, prevalence=None, *, prevalence_names=None, iters=30, convergence_tol=1e-05, kappa_estimation=None, kappa_ridge=0.001, em_tol=None, covariates=None, keep_eta_cov=True, reference='none', progress=None)

SAGE(num_topics, *, alpha=0.1, prior='laplace', prior_variance=1.0, optimize_interval=50, burn_in=200, seed=13, lbfgs_iters=20)
    Sparse additive generative model: the same topic worded differently across groups.
    .fit(data, groups, *, group_names=None, iters=1000, num_samples=5, sample_interval=25, progress=None, progress_interval=50, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10)

DMR(num_topics, *, beta=0.01, optimize_interval=50, burn_in=200, seed=13, alpha=0.1, prior_variance=1.0, alpha_epsilon=1e-10, lbfgs_iters=20, sampler='sparse', num_threads=1)
    Dirichlet-multinomial regression: a document-metadata prior on topic proportions.
    .fit(data, features=None, *, feature_names=None, iters=1000, num_samples=5, sample_interval=25, progress=None, progress_interval=50, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10, covariates=None, offset=None, num_threads=None)

GDMR(num_topics: 'int', *, degrees: 'list[int]', beta: 'float' = 0.01, optimize_interval: 'int' = 50, burn_in: 'int' = 200, seed: 'int' = 13, sigma: 'float' = 1.0, sigma0: 'float' = 3.0, decay: 'float' = 0.0, alpha: 'float' = 0.1, metadata_range: 'list[tuple[float, float]] | None' = None, lbfgs_iters: 'int' = 20, sampler: 'str' = 'sparse', num_threads: 'int' = 1) -> 'None'
    Generalized DMR with a smooth (Legendre-basis) prior over continuous covariates.
    .fit(data: "'Corpus | Sequence[Sequence[str]]'", features=None, *, metadata_names=None, iters: 'int' = 1000, num_samples: 'int' = 5, sample_interval: 'int' = 25, keep_theta_draws: 'bool' = True, convergence_tol: 'float' = 0.0, check_every: 'int' = 10, covariates=None, metadata=None, progress=None) -> 'None'

Scholar(num_topics, *, covariates=None, covariate_names=None, content=None, content_names=None, interactions=False, alpha=1.0, hidden_size=100, dropout=0.2, batch_size=200, lr=0.002, l2_prior_reg=0.0, l1_content_reg=0.0, convergence_tol=0.0, seed=13)
    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.
    .fit(data, *, covariates=None, labels=None, content=None, iters=None, convergence_tol=None, progress=None)

RTM(num_topics, *, link=None, inference='variational', alpha=None, beta=0.1, rho=None, negative_ratio=1.0, ridge=1.0, seed=13)
    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.
    .fit(data, links, *, iters=50, e_sweeps=3, e_inner=5, progress=None)

FactorialLDA(factor_sizes, *, sigma_alpha=1.0, sigma_alpha_bias=1.0, sigma_omega=0.5, sigma_omega_bias=10.0, delta0=0.1, delta1=0.1, alpha_bias_init=Ellipsis, omega_bias_init=Ellipsis, step_alpha_doc=0.01, step_alpha_corpus=None, step_alpha_bias=None, step_omega=0.001, step_omega_bias=None, step_beta=0.001, block_freq=1, weight_burnin=100, word_priors=True, sparsity=True, symmetric_word_prior=False, seed=13)
    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.
    .fit(data, *, iters=2000, samples=100, eval_every=0, omega_priors=None, observed_factors=None, progress=None)

AuthorTopic(num_topics, *, alpha=None, beta=0.01, seed=13)
    Author-Topic Model: each author has a topic distribution; documents mix their authors. Answers what an author writes about.
    .fit(data, authors, *, iters=1000, progress=None)

AuthorRecipientTopic(num_topics: 'int', *, alpha: 'float | None' = None, beta: 'float | None' = 0.1, seed: 'int' = 13)
    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.
    .fit(docs: 'Sequence[Sequence[str]]', *, authors: 'Sequence', recipients: 'Sequence[Sequence]', iters: 'int' = 1000, progress=None)

### Guided & supervised

KeyATM(keywords, *, num_topics=None, alpha=None, beta=0.01, beta_keyword=0.1, gamma1=1.0, gamma2=1.0, seed=13, estimate_alpha=True, sampler='sparse', num_threads=1)
    Keyword-assisted topics: anchor named topics with a few seed words each.
    .fit(data, *, iters=1500, covariates=None, feature_names=None, times=None, timestamps=None, num_states=5, weights='information-theory', num_threads=None, optimize_interval=50, burn_in=200, prior_variance=1.0, lbfgs_iters=20, progress_interval=0, prior_offset=None, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, report_interval=None, turbo_alpha_stride=1, progress=None)

SeededLDA(seed_words, *, residual=0, alpha=0.5, beta=0.1, weight=0.01, seed=13, seed_prior='frequency', sampler='sparse', seed_match='fixed', case_insensitive=False, num_threads=1)
    Seeded LDA: steer named topics toward supplied seed words.
    .fit(data, *, iters=2000, doc_topic_prior=None, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10, num_threads=None, progress=None)

GuidedNMF(num_topics, seed_words, *, guidance=3.0, lam=None, seed_weight=1.0, init='random', weighting='tfidf', convergence_tol=0.0, seed_match='fixed', case_insensitive=False, init_a=None, init_s=None, init_b=None, seed=13)
    Guided NMF: seed-word-guided semi-supervised NMF; the matrix-factorization analogue of SeededLDA.
    .fit(data, *, iters=None, convergence_tol=None, num_threads=None)

LabeledLDA(*, alpha=0.1, beta=0.01, seed=13, sampler='sparse', num_threads=1)
    Labeled LDA: each document label is a topic; tokens are restricted to its labels.
    .fit(data, labels, *, label_names=None, iters=1000, num_samples=5, sample_interval=25, progress=None, progress_interval=50, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10, num_threads=None)

SupervisedLDA(num_topics, *, alpha=0.1, seed=13, inference='variational')
    Supervised LDA: topics shaped to predict a per-document real-valued response.
    .fit(data, y, *, iters=25, var_iters=15, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=1, num_threads=None, progress=None)

DiscLDA(k_class, k_shared, *, alpha=None, beta=0.01, iters=1000, infer_sweeps=100, class_prior=None, seed=13)
    Discriminative LDA (Lacoste-Julien et al. 2008): topics split into per-class and shared blocks; reads how classes talk differently.
    .fit(data, y, *, iters=None, progress=None)

### Short text

GSDMM(num_topics, *, alpha=0.1, beta=0.1, seed=13, num_threads=1)
    Gibbs-sampling Dirichlet mixture: one topic per short document.
    .fit(data, *, iters=30, progress_interval=0, report_interval=None, num_threads=1, verbose=False, progress=None)

PT(num_topics, *, num_pseudo=100, alpha=0.1, beta=0.01, pseudo_doc_prior=0.1, seed=13, num_threads=1)
    Pseudo-document topic model: pool short texts into pseudo-documents.
    .fit(data, *, iters=1000, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10, num_threads=None, progress=None)

BTM(num_topics, *, alpha=None, beta=0.01, iters=1000, window=15, background=False, seed=13, num_threads=1)
    Biterm topic model: learns topics from corpus-level word co-occurrence (biterms).
    .fit(data, *, iters=None, num_threads=None, progress=None)

### Dynamic & hierarchical

DTM(num_topics, *, alpha=0.01, chain_variance=0.005, obs_variance=0.5, seed=13, init='random')
    Dynamic topic model: a fixed topic set whose word distributions drift across time slices.
    .fit(data, times, *, iters=20, progress=None)

DETM(num_topics, *, delta=0.005, hidden_size=800, eta_hidden_size=200, eta_nlayers=3, batch_size=1000, lr=0.005, wdecay=1.2e-06, grad_clip=None, convergence_tol=0.0, seed=13)
    Dynamic embedded topic model: embedding-factored topics that drift across time slices, fit as an amortized VAE.
    .fit(data, word_embeddings, vocabulary, *, times=None, timestamps=None, iters=100, convergence_tol=None, progress=None)

TopicsOverTime(num_topics, *, alpha=None, beta=0.1, seed=13)
    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).
    .fit(data, times=None, *, timestamps=None, iters=1000, progress=None)

HLDA(*, depth=3, gamma=1.0, beta=0.01, alpha=None, level_prior='dirichlet', gem_mean=0.5, gem_scale=100.0, seed=13, eta=None)
    Hierarchical LDA (nested CRP): a learned tree of super- and sub-topics.
    .fit(data, *, iters=500, num_threads=1, progress=None)

PA(num_super, num_sub, *, alpha=0.1, beta=0.01, seed=13, num_threads=1)
    Pachinko allocation: a DAG of super- and sub-topics.
    .fit(data, *, iters=1000, keep_theta_draws=True, num_theta_draws=25, convergence_tol=0.0, check_every=10, num_threads=None, progress=None)

### Embedding-based

KeyNMF(num_topics, *, top_n=25, metric=Ellipsis, seed=13)
    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.
    .fit(data, doc_embeddings, *, word_embeddings, vocabulary, iters=None, convergence_tol=None)

BERTopic(*, n_components=5, min_cluster_size=15, min_samples=None, nr_topics=None, window=4, stride=1, reducer='umap', n_neighbors=15, bm25=False, reduce_frequent=False, weighting='c-tf-idf', min_similarity=0.0, clusterer='hdbscan', num_clusters=None, resolution=1.0, knn_neighbors=15, diagnostics=True, min_dist=0.0, spread=1.0, n_epochs=0, negative_sample_rate=5, repulsion_strength=1.0, metric='cosine', seed=13)
    Cluster document embeddings; label topics by class-based TF-IDF.
    .fit(data, doc_embeddings)

Top2Vec(*, n_components=5, min_cluster_size=15, min_samples=None, reducer='umap', n_neighbors=15, clusterer='hdbscan', num_clusters=None, resolution=1.0, knn_neighbors=15, diagnostics=True, min_dist=0.0, spread=1.0, n_epochs=0, negative_sample_rate=5, repulsion_strength=1.0, metric='cosine', seed=13)
    Topics as dense regions in a joint document-word embedding space.
    .fit(data, doc_embeddings, *, word_embeddings=None, vocabulary=None)

SemanticSignalSeparation(num_topics, *, feature_importance=Ellipsis, iters=200, convergence_tol=0.0001, seed=13)
    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.
    .fit(data, doc_embeddings, vocab_embeddings, *, vocabulary=None)

ETM(num_topics, *, inference='em', convergence_tol=0.0001, sigma_shrink=0.0, prior_variance=1000000.0, max_inner=25, hidden_size=800, batch_size=1000, lr=0.005, wdecay=1.2e-06, seed=13, prior=Ellipsis, contrastive=False, contrastive_weight=0.5, contrastive_temp=0.5, em_tol=None)
    Embedded topic model: topic-word distributions factored through word embeddings.
    .fit(data, word_embeddings, vocabulary, *, iters=None, convergence_tol=None, progress=None)

GaussianLDA(num_topics, *, alpha=None, kappa=0.1, nu=None, psi_scale=3.0, init=Ellipsis, seed=13)
    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.
    .fit(data, word_embeddings, vocabulary, *, iters=None, progress=None)

FASTopic(num_topics, *, lr=0.002, dt_alpha=3.0, tw_alpha=2.0, theta_temp=1.0, convergence_tol=1e-06, sinkhorn_iters=5000, sinkhorn_tol=0.005, seed=13, em_tol=None)
    Topics from optimal-transport plans between document, topic, and word embeddings.
    .fit(data, doc_embeddings, *, iters=None, convergence_tol=None, progress=None)

CombinedTM(num_topics, *, alpha=1.0, hidden_size=100, dropout=0.2, batch_size=200, lr=0.002, convergence_tol=0.0, seed=13, prior=Ellipsis, contrastive=False, contrastive_weight=0.5, contrastive_temp=0.5)
    Contextualized ProdLDA: encoder reads the bag of words plus a document embedding.
    .fit(data, doc_embeddings, *, iters=None, convergence_tol=None)

ZeroShotTM(num_topics, *, alpha=1.0, hidden_size=100, dropout=0.2, batch_size=200, lr=0.002, convergence_tol=0.0, seed=13, prior=Ellipsis, contrastive=False, contrastive_weight=0.5, contrastive_temp=0.5)
    Contextualized ProdLDA: encoder reads the document embedding alone, enabling cross-lingual transfer.
    .fit(data, doc_embeddings, *, iters=None, convergence_tol=None)

InfoCTM(num_topics, *, mi_weight=30.0, mi_temperature=0.2, pos_threshold=0.4, hidden_size=100, dropout=0.0, lr=0.002, convergence_tol=0.0, seed=13, languages=None)
    Cross-lingual: two ProdLDA models aligned by a bilingual dictionary through a mutual-information term.
    .fit(data_a, data_b, *, dictionary, embeddings_a=None, embeddings_b=None, iters=None, batch_size=128, progress=None)

### Ideal point

Wordfish(*, beta_prior_sd=3.0, theta_prior_sd=1.0, min_count=1, convergence_tol=1e-06, seed=13)
    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.
    .fit(data, *, group=None, control=None, anchors=None, iters=None, convergence_tol=None, progress=None)

Wordshoal(*, theta_prior_sd=1.0, loading_prior_sd=0.5, intercept_prior_sd=0.5, tau_prior=1.0, min_count=1, convergence_tol=0.001, seed=13)
    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.
    .fit(data, *, speakers, domains, anchors=None, iters=None, convergence_tol=None)

TBIP(num_topics, *, a_gamma=0.3, b_gamma=0.3, iters=7000, batch_size=512, learning_rate=0.05, min_count=1, seed=13)
    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.
    .fit(data, *, group=None, iters=None, batch_size=None, learning_rate=None, progress=None)

PartyEmbeddings(num_dims=2, *, vector_size=200, window=20, min_count=5, negative=5, sample=0.0001, learning_rate=0.025, seed=13)
    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.
    .fit(data, *, group, control=None, anchors=None, iters=5)

### LLM-based

TopicGPT(*, backend: 'Optional[Callable[[str], str]]' = None, model: 'Optional[str]' = None, hierarchical: 'bool' = False, assignment: 'str' = 'hard', sample: 'Optional[int]' = None, max_topics: 'Optional[int]' = None, min_topic_count: 'int' = 1, temperature: 'float' = 0.0, seed: 'int' = 13, prompts: 'Optional[dict]' = None) -> 'None'
    LLM-driven topic discovery: prompt a model to propose, refine, and assign a topic taxonomy with descriptions.
    .fit(data, *, metadata=None) -> "'TopicGPT'"

Experimental models are omitted; enable_experimental() then list_models(experimental=True) to see them.