mini-jev-minicpm5-2b-shell-safety-phase1b-onnx

Mini-Jev — shell-safety Phase-1b (ONNX)

Browser-ready allow / warn / block classifier for shell commands. Closed-set scoring over candidate label tokens (one forward pass; no text generation).

Built for the Mini-Jev WebGPU demo.

WebGPU on Apple Chrome (Metal)

On Apple M-series Chrome with WebGPU / ORT Metal, prefer q8:

• Prefer dtype: "q8" → onnx/model_quantized.onnx (+ model_quantized.onnx_data, ~2.35 GB). Alias: onnx/model_q8.onnx.
• q8 uses IntegerOps (MatMulInteger / MatMulConstBOnly) — no MatMulNBits, so it avoids Metal unaligned accesses.
• fp16 (~5.0 GB with model_fp16.onnx_data) can RangeError: Out of memory in Chrome/Transformers.js on M2 Max.
• q4 / q4f16 may fail on Metal with MatMulNBits unaligned accesses.

WASM can still use q4 for the smaller download.

What's in this repo

Transformers.js layout for @huggingface/transformers AutoModelForCausalLM:

• config.json, tokenizer files, quantize_config.json
• onnx/model.onnx (+ _data) — fp32 baseline (~5.4 GB data)
• onnx/model_quantized.onnx (+ _data) — q8 / IntegerOps; recommended for Apple Chrome WebGPU (~2.35 GB). Alias onnx/model_q8.onnx
• onnx/model_fp16.onnx (+ _data) — all-fp16 (~5.0 GB); kept for non-browser / when heap allows
• onnx/model_q4.onnx — ~1.8 GB; may fail on Metal WebGPU (MatMulNBits)
• onnx/model_q4f16.onnx — optional; may fail on Metal WebGPU

Load

import { AutoModelForCausalLM, AutoTokenizer } from "@huggingface/transformers";

const model_id = "nicolasembleton/mini-jev-minicpm5-2b-shell-safety-phase1b-onnx";
const tokenizer = await AutoTokenizer.from_pretrained(model_id);
const model = await AutoModelForCausalLM.from_pretrained(model_id, {
  dtype: "q8", // IntegerOps; use "q4" on WASM if you need smaller
  device: "webgpu", // or "wasm"
});

Model

Training data (summary)

Ternary gold was planted after remapping a public binary shell-risk set; binary labels were not treated as ground truth.

ShellRisk upstream slices used in train (approx.): swesmith, terminalbench, nl2bash, atomicredteam, capped gtfobins. Prefer upstream licenses; this card does not relicense them.

Frozen smoke metrics (PyTorch Phase-1b)

Planted gold on the 105-item smoke panel (not a closed-menu score):

Browser ONNX / WebGPU may differ slightly from these PyTorch numbers.

Intended use

Research and demos of in-browser agent tooling gates. Not a substitute for sandboxing, allowlists, or human review on real systems.