Zero-Shot Text Classifier — Free, Server-Side, No API Key

The page implements the entailment-as-zero-shot-classification recipe first published by Yin, Hay & Roth (2019). The same recipe is the default behind Hugging Face's zero-shot-classification task — and that is the API we call server-side from this page.

1. Build a hypothesis per label

Your text is treated as the premise. For each candidate label L, the tool substitutes the label into the hypothesis template — the default is This text is about {}.. With label urgent, the hypothesis becomes “This text is about urgent.” A purpose-specific template (“This support ticket is {}.”) typically raises accuracy, which is why the field is editable in the Advanced disclosure.

2. Run an NLI head

The (premise, hypothesis) pair is tokenised and passed through valhalla/distilbart-mnli-12-3, a distilled BART encoder-decoder fine-tuned on MultiNLI (Williams, Nangia & Bowman, 2018). The classification head emits three logits per pair: contradiction, neutral, and entailment. The pipeline discards neutral and applies softmax to [contradiction, entailment] to produce the entailment probability — the chance that the hypothesis is entailed by the premise.

3. Aggregate across labels

In single-label mode the per-label entailment probabilities are softmaxed across all labels, so the scores sum to 100 %. In multi-labelmode each label's entailment probability is reported as-is — multiple labels can be high. Pick single-label when the categories are mutually exclusive (news topic, primary intent), multi-label when they overlap (“urgent” AND “billing-related”).

for each label L_i:
  hypothesis_i  = template.replace("{}", L_i)
  [c, n, e]     = NLI(premise = text, hypothesis = hypothesis_i)
  entail_i      = softmax([c, e])[1]      // discard neutral

single-label:  scores = softmax([entail_1, ..., entail_N])
multi-label:   scores = [entail_1, ..., entail_N]

4. Server-side inference & privacy

The model runs server-side on the Hugging Face Inference API. Your text and labels are POSTed to /api/tools/zero-shot, forwarded once to Hugging Face, and the response is returned to your browser. Nothing is written to disk or any database; the route uses cache: "no-store" so there is no edge cache either. Inputs are capped at 4,000 characters and 12 labels to keep latency predictable. A deterministic word-overlap baseline runs alongside the model on every request as a transparent cross-check — when both rankings pick the same top label, the confidence is high; when they disagree, the page surfaces that explicitly.

• Yin, Hay & Roth (2019) — Benchmarking Zero-shot Text Classification (arXiv)
• Hugging Face — Zero-Shot Classification task page (template and multi_label spec)
• Lewis et al. (2019) — BART: Denoising Sequence-to-Sequence Pre-training (arXiv)
• Williams, Nangia & Bowman (2018) — MultiNLI corpus (ACL Anthology)
• Hugging Face — valhalla/distilbart-mnli-12-3 (model card, distillation details)
• Hugging Face Inference API documentation

Model card, Inference API endpoint, and method citations were last cross-checked on 2026-05-12. The page is reviewed each time the upstream model repository is updated or a new MNLI-based NLI checkpoint becomes the recommended default.