gpu-ask.js (v13)

On-device agent-question tagger for the browser. Two-seed ensemble (~571k parameters × 2), distilled from an Ettin-150m teacher. Emissions are ONNX; features, linear-chain CRF, and the ask-tool compiler ship as TypeScript in @ai-ecoverse/gpu-ask.js.

Live demo · npm · GitHub

Files (`v13/`)

File	Role
`member0.onnx`, `member1.onnx`	Emission networks (ensemble average at runtime)
`config.json`	Feature / label / member layout
`crf.json`	Start and transition scores for Viterbi
`static_vocab.json`	Static potion-32 table
`manifest.json`	SHA-256 checksums

Usage

import * as ort from "onnxruntime-web/wasm";
import { loadAsk, parse } from "@ai-ecoverse/gpu-ask.js";

const ask = await loadAsk(
  "https://huggingface.co/ai-ecoverse/gpu-ask.js/resolve/main/v13",
  { ort },
);
const { questions } = await parse(ask, agentMessage);

Weights are also bundled under models/v13/ in the npm package (bundledModelUrl()).

Metrics

Pooled out-of-fold on 4,930 hand-labeled public agent turns (v13, constrained kind rerank in Python): false questions 3.7%, kind 85.4%, options_ok 67.5%. The browser path uses compiler kind (no Python-only rerank).

Training

See ai-ecoverse/gpu-questions (docs/training.md, recipe v13). Architecture follows gpu-time.

License

Apache-2.0. Part of AI Ecoverse.