--- title: "Zero-Shot Classifier - Private Text Labels" description: "Use a private zero-shot classifier to rank English text against labels you choose. Runs on-device with WebGPU or a CPU fallback, with no text upload." url: https://candoya.com/tools/zero-shot-classifier/ locale: en type: browser-tool --- # Free Zero-Shot Classifier Category: [Social & Creator](https://candoya.com/tools/social/) ## What is a zero-shot classifier? A zero-shot classifier sorts text into labels you provide without needing labeled examples or model training. This tool turns each label into a natural language inference test, ranks the matches, and runs the English MobileBERT model inside your browser. Your text stays on your device while the public model downloads separately. **This classifier uses Xenova's MobileBERT MNLI model through Transformers.js.** It compares your text with a simple hypothesis for every candidate label. Choose one mode makes all labels compete and normalizes their scores across the set. Allow several mode treats each label independently, which suits passages that can belong to multiple categories. The WebGPU path uses quantized model weights of about 21 MB. The WebAssembly CPU fallback uses about 26 MB, plus tokenizer files and the runtime. Browsers normally cache downloads, but private browsing, cleared storage, or storage pressure can cause another download. The text you enter is passed only to the browser worker, not to the model host. ## How to use it 1. **Enter text and candidate labels** - Paste one focused English passage, then add between 2 and 12 short labels separated by commas or new lines. 2. **Choose how labels relate** - Use Choose one when exactly one category should win. Use Allow several when several labels can all describe the passage. 3. **Run and review the ranking** - Start classification, allow the first model download to finish, and compare the full score list rather than relying on the leading label alone. 4. **Refine or copy** - Rewrite vague or overlapping labels and run again. Copy the ranked scores when the label set gives a useful comparison. ## Who it's for - **Support teams** testing labels such as billing, bug report, feature request, and praise before designing a routing workflow. - **Product researchers** exploring themes in English survey answers or interview notes without uploading unpublished text for inference. - **Meeting and voice-note users** comparing a transcript excerpt with labels such as decision, action item, risk, or follow-up. - **Developers and analysts** prototyping a label set before collecting examples for a supervised classifier or adding human review rules. ## Expert note Zero-shot scores depend on label wording and on the other labels in the set. In Choose one mode, adding a near-duplicate label can split the normalized score; in Allow several mode, two overlapping labels can both score highly. Test a label set on known examples before using it for real routing, and keep a person in the loop for consequential decisions. ## Frequently asked questions ### Is my text uploaded to a server? No. Classification runs in a worker inside your browser, and your pasted text is not sent to CanDoYa, jsDelivr, or Hugging Face. The worker does request the public Transformers.js runtime and model files. Those download requests do not include the text or labels you entered. ### Is the zero-shot classifier free? Yes. There is no account, API key, payment, or server-side usage quota. Your device performs inference. The practical costs are the first model download, browser storage, memory, processing time, and any data charge from your network provider for downloading the public files. ### How much text and how many labels can I use? The tool accepts 2 to 12 unique labels and classifies the first 1,600 Unicode characters of the normalized text. Each label can contain up to 60 characters. Longer documents should be split into coherent passages because one score list cannot describe every topic in a mixed document. ### What is the difference between Choose one and Allow several? Choose one assumes the labels are mutually exclusive, so their scores are normalized together and total about 100 percent. Allow several compares entailment and contradiction separately for every label. Its scores are independent, so several labels can rank highly and the total does not need to equal 100 percent. ### What do zero-shot classification scores mean? Scores express how strongly this English NLI model matches the text to the supplied labels under the selected mode. They are useful for ranking and comparison, but they are not calibrated guarantees that a label is correct. Domain language, label wording, ambiguity, and out-of-distribution text can change them. ### Which languages does the classifier support? The selected uncased MobileBERT MNLI model was trained and fine-tuned for English, so this page marks the tool English only. Other languages may still produce numbers, but those values lack the same training basis and should not be treated as reliable classification results. ### Does it work without WebGPU? Yes. The worker tries WebGPU when the browser and graphics adapter expose it. If WebGPU is unavailable or fails, the tool restarts with the WebAssembly CPU backend. The fallback uses the same MobileBERT classifier and keeps text local, but model loading and inference can take longer. ### Can I use the result for automatic decisions? Use it for experiments, drafting label sets, or low-risk triage with human review. Do not use one model score by itself for employment, credit, health, safety, legal, moderation, or other consequential decisions. Validate the label set with representative examples and define a review path for uncertain cases. ## Related tools - [Sentiment Analysis](https://candoya.com/tools/sentiment-analysis/) - [Keyword Extractor](https://candoya.com/tools/keyword-extractor/) - [Named Entity Recognition](https://candoya.com/tools/named-entity-recognition/) - [Language Detector](https://candoya.com/tools/language-detector/) - [Text Similarity](https://candoya.com/tools/text-similarity/) --- *Markdown edition of , published for AI agents and other automated readers. Index of key pages: .*