CoolFace
Apppublic

roskosmos19/Zero-shoot-clasification

sourceHugging Facemitupdated 2d agoView on Hugging Face
1likes
App README

Zero-Shot Decisions · Browser

Fully client-side zero-shot classification powered by Transformers.js (WebGPU / WASM). No server, no ZeroGPU, no API keys. Enter a compatible Hugging Face model and score options against a state.

What it does

  1. 1.Enter a Hugging Face model ID that supports zero-shot classification (or pick a suggestion).
  2. 2.Provide a state (context) and one or more questions with allowed options.
  3. 3.The model scores every option via the zero-shot-classification pipeline.
  4. 4.Results show calibrated-style probabilities and the top choice per question.

Requirements

  • —Modern browser (Chrome / Edge recommended for WebGPU; WASM fallback available).
  • —First load downloads the model weights (cached afterwards).

Suggested models

Model IDNotes
Xenova/mobilebert-uncased-mnliSmall, fast
Xenova/nli-deberta-v3-xsmallStrong small NLI
Xenova/bart-large-mnliLarger, higher quality
MoritzLaurer/deberta-v3-base-zeroshot-v1.1-all-33High-quality zero-shot (if ONNX available)

Only models that Transformers.js can load (ONNX / converted) will work. Raw PyTorch checkpoints are not supported without conversion.

Local development

bash
python -m http.server 8000

Open http://localhost:8000 in a modern browser.