nodd

Try it

Runs entirely in your browser. The first run downloads the selected model from Hugging Face; your text never leaves this page.

Below the calibrated threshold a decision should be escalated to a larger model. Models are trained on synthetic data; expect lower accuracy on real text.

How it works

task spec (YAML) → collect inputs → teacher labels
  → train → calibrate → evaluate → export → browser

A small MiniLM sentence encoder is fine-tuned with a classification head, exported to ONNX q8 and run with transformers.js. Every base model that fits the download budget is trained and the best fit is kept. Each prediction is a typed decision with a calibrated confidence and a threshold chosen for a target precision. Exports are checked for parity: the browser must match Python on the test set.

Released models

Test macro-F1

q8 download, MB

q8 test F1 on small synthetic holdouts (50–257 cases); real-world accuracy is unmeasured. Chromium WASM and WebGPU predictions agree with Python ONNX on 99.6–100% of test cases. Training report

Picking a base: comment moderation

“Handled” is the share of test cases answered without escalation; p95 is browser latency.

Use it

uv run nodd run examples/comment_moderation.yaml

import { nodd } from "@nodd/browser";
const m = await nodd.load("/models/comment_moderation/v4");
const d = await m.decide("Buy cheap followers at …");
// { label: "spam", confidence: 0.99, … }