# Train a local Naive Bayes text classifier Train a naive Bayes classifier from labeled examples, then classify text with multinomial, Bernoulli, or complement scoring and token explanations. ## Run it - **CLI:** `gizza tool naive-bayes-text-classifier "spam,win a free prize now spam,free money click here ham,meeting at ten tomorrow ham,lunch with the team" 'text=claim your free money now'` - **Web:** https://gizza.ai/tools/naive-bayes-text-classifier/ - **Agents:** machine-readable descriptor (parameters JSON Schema) at https://gizza.ai/tools/naive-bayes-text-classifier/tool.json ## Inputs - `training_data` — Training data _(field)_ - `text` — Text to classify _(field)_ - `separator` — Training separator _(field)_ - `input_mode` — Classify as _(field)_ - `model` — Naive Bayes variant _(field)_ - `alpha` — Smoothing alpha _(field)_ - `ngram_max` — Max n-gram length _(field)_ - `lowercase` — Lowercase before tokenizing _(field)_ - `remove_stopwords` — Remove English stop words _(field)_ - `min_count` — Minimum token count _(field)_ - `priors` — Class priors _(field)_ - `top_k` — Classes to show _(field)_ - `explain` — Explain top tokens _(field)_ - `output` — Output format _(field)_ ## Output - Classification result (text) ## Query parameters Open the tool pre-filled and auto-run via URL: - `training_data` — Training data - `text` — Text to classify - `separator` — Training separator - `input_mode` — Classify as - `model` — Naive Bayes variant - `alpha` — Smoothing alpha - `ngram_max` — Max n-gram length - `lowercase` — Lowercase before tokenizing - `remove_stopwords` — Remove English stop words - `min_count` — Minimum token count - `priors` — Class priors - `top_k` — Classes to show - `explain` — Explain top tokens - `output` — Output format Example: `https://gizza.ai/tools/naive-bayes-text-classifier/?training_data=spam%2Cwin%20a%20free%20prize%20now%0Aspam%2Cfree%20money%20click%20here%0Aham%2Cmeeting%20at%20ten%20tomorrow%0Aham%2Clunch%20with%20the%20team&text=claim%20your%20free%20money%20now&separator=auto&input_mode=single&model=multinomial&alpha=1&ngram_max=1&lowercase=true&remove_stopwords=true&min_count=1&priors=empirical&top_k=3&explain=true&output=report` --- ## About this tool Naive Bayes Text Classifier trains a small supervised text model from the examples you paste, then immediately classifies new text in the browser or CLI. Each training line is `labeltext`, such as `spam,win a free prize now`. The separator can be auto-detected from tab, comma, pipe, or colon, or forced when your examples need it. The default multinomial model is the classic bag-of-words baseline for spam, topic, ticket, and short-message classification. Bernoulli mode uses token presence/absence, and complement mode is useful when one label has many more examples than another. Controls cover smoothing alpha, n-grams, case folding, English stop-word removal, rare-token pruning, class priors, class-list length, explanations, and report vs JSON output. ### Worked example Training data: ```text spam,win a free prize now spam,free money click here spam,claim your free gift ham,meeting at ten tomorrow ham,lunch with the team ham,project update attached ``` Text to classify: ```text claim your free money now ``` With default settings, the model predicts `spam`, lists the class probabilities, and shows the tokens that pushed the decision toward spam over the runner-up. Switch `input_mode=lines` to classify each non-blank input line as a separate document. ### Limits and edge cases - Training data is capped at 1 MiB, 20,000 examples, 200 labels, and 200,000 vocabulary tokens after `min_count` filtering. - Text to classify is capped at 256 KiB; batch mode accepts up to 1,000 non-blank lines. - `ngram_max` is limited to 3 so vocabulary growth stays predictable in WASM. - A single label is rejected; the model needs at least two distinct labels. - Unseen input tokens are ignored. If none of the input tokens appeared in training, the prediction comes from class priors alone and the report says so. - This is a stateless local baseline, not a persisted ML project. It does not save models, run validation splits, or download pre-trained language models. ## FAQ
How should I format the training data? Use one labeled example per line: `label,text`. The first separator on the line splits the label from the example text, so the text may contain more commas after that. Tabs, pipes, and colons are also supported; choose the separator explicitly if auto-detection guesses wrong.
Which model variant should I choose? Start with `multinomial` for ordinary word-count text classification. Use `bernoulli` for very short texts where word presence matters more than repeated words. Try `complement` when the labels are imbalanced and the largest class otherwise dominates too easily.
What does alpha smoothing do? `alpha` adds a small count to every token/class pair so a word that was unseen for one label does not make that label impossible. The default `1` is Laplace smoothing. Lower values make the classifier more confident in the pasted examples; higher values flatten the probabilities.
Can I reuse or export the trained model? No. The tool trains from scratch on every run and returns only the prediction/report. That keeps the tool stateless and local. If you need persisted models, cross-validation, or deployment, use a dedicated machine-learning workflow after validating the idea here.
## Related tools - [RAKE Keyword Extractor](https://gizza.ai/tools/rake-keywords/): Extract the top keywords and keyphrases from any document using the RAKE algorithm — ranked by relevance, right in your browser. Nothing is uploaded. - [ARFF Converter](https://gizza.ai/tools/arff-converter/): Convert Weka ARFF datasets to CSV and CSV tables back to ARFF locally — nominal attributes, numeric types, dates, sparse rows, missing values, and type rows. - [Regression model trainer](https://gizza.ai/tools/regression-model-trainer/): Fit linear, ridge, or bounded random-forest regression from a pasted numeric table and report R², RMSE, MAE, coefficients, or feature importance. - [Z-Score Normalize](https://gizza.ai/tools/z-score-normalize/): Standardize a list of numbers with z-score, min-max, max-abs, or robust median/IQR scaling. Population or sample standard deviation, private and browser-local. - [Extract Action Items from Meeting Notes](https://gizza.ai/tools/action-item-extractor/): Extract action items, owners, and decisions from meeting notes or daily notes with deterministic rules. Markdown checklist or JSON, private in-browser.