DistilBERT SST-2 — read the sentiment as you type

Why it matters

Sentiment is the entry point to understanding text at scale — and doing it in the browser means you can react to how a user feels as they type, without a round-trip and without their words ever leaving the device. That unlocks fast, private, always-available text intelligence.

• Live feedback & review triage — route angry messages to a human instantly, before submit.
• Private moderation — flag hostile drafts client-side; no PII sent to a server.
• Tone-aware writing aids — nudge an email that reads harsher than intended.
• Offline analytics — score a whole survey export in the tab, no upload.

Use cases

See also the MiniLM multi-model demo, which ranks search hits with embeddings and then labels each hit's tone with this model.

How the API works

Classification is one call; topk returns both class scores:

import { pipeline } from "@huggingface/transformers";

const classify = await pipeline(
  "text-classification",
  "Xenova/distilbert-base-uncased-finetuned-sst-2-english",
); // downloads + caches ~65 MB once

const out = await classify("best coffee all year", { topk: 2 });
// → [{ label: "POSITIVE", score: 0.9998 }, { label: "NEGATIVE", score: 0.0002 }]

"See inside" re-uses that same call: it classifies the full text plus one copy per word with that word deleted (batched in a single pass), then subtracts the POSITIVE log-odds to attribute the shift to each word. All of it runs in a Web Worker so typing never janks.

References

• Model card — DistilBERT SST-2 (Xenova ONNX)
• Paper — DistilBERT (Sanh et al., 2019, arXiv:1910.01108)
• SST — the Stanford Sentiment Treebank
• Transformers.js documentation