Dom: hotdog-rain training checkpoint
Richard Jones's experimental portfolio assistant, fine-tuned from
Qwen2.5-0.5B-Instruct using LoRA. This release preserves the trained adapter from
run dom-ui-20260804123828-a9e36621.
Artifact status
The adapter/ directory contains the trained PEFT weights and tokenizer.
The root tokenizer/config and onnx/model_q4.onnx contain the merged, 4-bit
browser-ready checkpoint. The ONNX file is approximately 770 MB; loading is
explicitly opt-in and memory-constrained devices may not support it.
Observed evaluation
The saved pre-export evaluation passed both of these requests:
Make it rain hotdogs over the portfolio.produced a PAGE_ACTION_JSON call towindow.__shoegunDomEffects?.hotdogRain().Stop the hotdog rain.produced a PAGE_ACTION_JSON call towindow.__shoegunDomEffects?.stopHotdogRain().
The full saved evaluation passed 7 of 29 checks (24.1%). This is an experimental checkpoint, not a reliable general-purpose assistant. These results are from the original adapter evaluation, not an ONNX or browser evaluation.
The merged q4 export separately passed both hotdog-rain requests in
Transformers.js CPU inference on September 7, 2026, producing the correct
hotdogRain() and stopHotdogRain() calls. This checks learned output directly,
without the portfolio's deterministic command shortcut. WebGPU performance and
general capabilities are not established by these two checks.
import { pipeline } from '@huggingface/transformers';
const dom = await pipeline('text-generation', 'Brazenle/dom-hotdog-rain', {
dtype: 'q4', device: 'wasm'
});
// Pin a published commit revision for a production deployment.
Host integration
The model proposes an action; the portfolio implements the animation and executes approved actions. Model weights do not themselves draw hotdogs. The host must validate actions and report actual execution outcomes. Do not execute arbitrary model-generated JavaScript in a privileged context.
Provenance and privacy
Base model: Qwen/Qwen2.5-0.5B-Instruct. LoRA rank 16, alpha 32, dropout 0.05. Only the selected model artifacts are published here. Raw training captures, private datasets, optimizer states, and local environment files are excluded.