How to use AI sentiment analyzer
- Enter an English review or comment within 500 characters, or load the English example. This does not start a model download.
- Select Analyze sentiment. First use downloads about 67.6 MB of quantized model weights plus tokenizer and runtime files; progress is shown for the current file.
- Wait for browser inference and review the model label and score with the original text. There are only positive and negative labels.
- Edit the passage to clear its old result. Editing during analysis cancels it, and Cancel analysis lets you stop downloading or processing.
A real pretrained classifier
The tool uses the ONNX conversion Xenova/distilbert-base-uncased-finetuned-sst-2-english, pinned to a specific repository revision, with q8 weights. Its upstream DistilBERT model was fine-tuned for English sentiment classification on SST-2. It performs model inference locally rather than applying a keyword list or calling a paid text-generation service.
The supported result labels are POSITIVE and NEGATIVE. Neutral or mixed opinions still receive one of those labels. A high model score describes this classifier's preference between its labels; it does not establish that the interpretation is correct for a particular person, context, or text.
Input and resource limits
Use a short English passage within 500 characters. The worker also checks the model's 512-token input limit before classification and rejects an overlong token sequence instead of silently truncating it. Other languages have not been validated for this English model.
The model runs on CPU through WebAssembly in a separate browser worker, using one WASM thread. This avoids requiring WebGPU, but download speed, memory, browser restrictions, and device performance still affect whether it can run. A failed download or inference produces an explicit error, not a substituted result.
What is downloaded and what stays local
No model download begins until you select Analyze sentiment. First use requests public model assets from Hugging Face and browser runtime assets from the library's CDN path. The q8 weight file alone is 67,581,197 bytes; tokenizer, configuration and runtime files add to the total.
Input text is passed only to the worker in this page for classification. The conversion code does not put it into a server request or store text history. The browser may cache the public model files for later use; external hosts can still observe normal download request metadata. Closing the page or cancelling stops the worker, while browser-cached files may remain.
Which format do you need?
| Your next task | Input or choice | Output or result |
|---|---|---|
| Classify a short passage | English text you enter | a positive or negative model label and score |
| Evaluate a neutral opinion | neutral or mixed wording | the binary model still chooses one label; review its limitation |
| Keep text off an inference server | text in this page's worker | local CPU/WASM inference; public assets download separately |
Questions about this tool
Is this actually a machine-learning model?
Yes. The page loads pinned, quantized DistilBERT ONNX weights and runs classification in a browser worker with Transformers.js. It does not generate a pretend result from keywords.
Will my text be uploaded?
The tool passes your text to its local browser worker, not a remote inference endpoint. Public model and runtime files are downloaded separately, and those hosts can observe normal request metadata.
How large is the first download?
The q8 weight file is 67,581,197 bytes, about 67.6 MB. Tokenizer, configuration and runtime files add to that amount. No model download starts until Analyze sentiment is selected.
What does the score mean?
It is the score assigned by this binary classifier to its selected label. It is not a guaranteed accuracy rate or proof of a person's intended emotion.
Does it support neutral text or other languages?
The model has only positive and negative labels. English is its intended language; other languages are not validated. Neutral, mixed or sarcastic text may be misclassified.
What if the model cannot load?
You receive a download or browser inference error and no result. Check your connection and browser resources, then retry. Editing or Cancel analysis stops the worker; processing is also stopped after a 3-minute timeout.