smollm2-pg-sft-browser-q4f16

SmolLM2 PG-style browser rewriter

This is the merged SmolLM2-1.7B SFT paragraph rewriter trained on the teacher rewrite corpus plus Paul Graham style anchors. It is packaged as an ONNX Runtime Web artifact so the model can be downloaded and executed in the user's browser with Transformers.js.

The weights were exported to ONNX FP16 and then converted to weight-only q4f16 MatMulNBits operators with block size 128. This is the browser runtime artifact; it is separate from the server-side AutoAWQ/GEMM checkpoint.

Browser usage

import { pipeline } from "https://cdn.jsdelivr.net/npm/@huggingface/[email protected]/+esm";

const rewriter = await pipeline(
  "text-generation",
  "matthewagi/smollm2-pg-sft-browser-q4f16",
  { device: "webgpu", dtype: "q4f16" },
);

The first load downloads roughly 1.1 GB and then the browser caches the model. WebGPU is recommended; a WASM CPU fallback is possible but substantially slower. No inference server is required.

The base model is released under its upstream license. The SFT data and adapter are project artifacts; consult the project repository for provenance.