saidthefox/systema-minion-0.6b-v4-ONNX
systema-minion-0.6b-v4 (ONNX / transformers.js)
ONNX build of the [Systema Constructum](https://quartermachines.website/evolution) judge — a LoRA fine-tune of Qwen/Qwen3-0.6B that decides whether a proposed taxonomy entry or edge should be ADVANCED or STRUCK, trained on the game's own court rulings. This is the in-browser build; the served weights and full story are in the GGUF repo.
onnx/model_quantized.onnx— int8 (per-channel), fordtype: "q8".
Use with transformers.js
import { AutoTokenizer, AutoModelForCausalLM } from "@huggingface/transformers";
const id = "saidthefox/systema-minion-0.6b-v4-ONNX";
const tok = await AutoTokenizer.from_pretrained(id);
const model = await AutoModelForCausalLM.from_pretrained(id, { dtype: "q8", device: "webgpu" });
const text = tok.apply_chat_template([
{ role: "system", content: "You are an agent in Systema Constructum - a taxonomy of everything HUMANS HAVE MADE (never natural kinds; entries over natural referents denote the made concept, Law 2a). Reply with EXACTLY ONE JSON object on ONE line." },
{ role: "user", content: 'Judge this pending EDGE.\nprivilege -DEPENDS_ON-> hierarchy\nfiler\'s note: Privilege requires a hierarchy to exist.\nReply: {"move":"judge","targetId":"cX","vote":"ADVANCE" or "STRIKE","reasoning":"..."}' },
], { tokenize: true, add_generation_prompt: true, enable_thinking: false, return_tensor: false });
const out = await model.generate({ ...tok(text), max_new_tokens: 200, do_sample: false });Note on WebGPU
Runs on the GPU under transformers.js on desktop Chrome/Edge and Android. On iOS Safari, transformers.js's WebGPU backend does not initialize, so it falls back to CPU/wasm (correct, but slow). For GPU inference on iPhone use an MLC/WebLLM build instead.
Evaluation & limitations
See the GGUF repo card: strike recall 0.626 vs a stock base that never strikes; narrow to the judging format; a 0.6B model that can reach a right verdict via shaky reasoning. Apache-2.0, inherited from Qwen3.