Qwen3-1.7B fine-tuned for the "Voice Assistant for Web" browser extension (q4f16_1, WebLLM)
A LoRA fine-tune of Qwen/Qwen3-1.7B for one narrow job:
the on-device assistant of a Chrome extension for hands-free browsing (for people with motor or
visual impairments). Given a description of the current page, open tabs and recent commands, it
replies with JSON — something to say aloud and up to three actions in a small command language
(scroll.down, click vaw-12, type …, search youtube: …).
Converted to MLC's q4f16_1 format for WebLLM; use it with the
standard compiled library for Qwen3-1.7B-q4f16_1 (v0_2_84). Not a general chat model.
Training
- Data: 2,860 synthetic examples in the extension's exact prompt format — answers about invented web pages, clicks by element reference, commands phrased many ways, typing into fields, tabs, searches, clarifying questions, and pages containing prompt-injection text that must be ignored. Natural language written by Gemini (free tier) as a teacher; actions labelled by rule.
- LoRA on 16 of 28 layers, 1e-4, batch 4, loss on replies only (MLX, Apple M4 Pro).
- Converted with a re-implementation of MLC's q4f16_1 quantisation verified to reproduce the official converted Qwen3-1.7B byte for byte.
Evaluation (40 held-out scenarios)
| Model | Benchmark (40 held-out scenarios) | Median latency |
|---|---|---|
| This model, 4-bit q4f16_1 in WebLLM (Chrome, WebGPU, Apple M4 Pro) | 32/40 (80%) | 1.4 s |
| Same fine-tune, full precision (MLX) | 35/40 (87.5%) | 0.36 s |
| Stock Qwen3-1.7B q4f16_1 in WebLLM | 12/40 (30%) | 1.5 s |
| Stock Qwen3-1.7B, full precision (MLX) | 24/40 (60%) | 0.45 s |
| Cloud reference: Groq gpt-oss-20b / Gemini 3.5 Flash-Lite | 37/40 (92.5%) | 0.45 s / 0.9 s |
Quantised with a per-group clipping search (the q4f16_1 format is unchanged); plain round-to-nearest scores 31/40 in WebLLM.
The scenarios' pages and sentences are not in the training data. Details and reproduction:
the extension's docs/EVALUATION-SUMMARY.md and training/README.md.
Limitations
Small model: it still chooses the wrong action more often than cloud models, can be wrong on general knowledge, and must not be trusted to act without the extension's own checks (which allow only known commands, real page elements and web addresses, and ask before risky actions).
- Downloads last month
- 16