qwen3-4b-journalist-ONNX

An investigative journalism model fine-tuned on Qwen/Qwen3-4B (4B params), converted to ONNX for in-browser inference via Transformers.js + WebGPU. Same training data and domain coverage as gemma4-e4b-journalist.

Built by Buried Signals for edge/browser inference in OSINT Navigator.

Usage

This model runs entirely in the browser. No server-side inference required.

In OSINT Navigator: Toggle "Run locally" in the search bar. The model downloads once (~2.5 GB) and runs via WebGPU.

With Transformers.js:

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

const tokenizer = await AutoTokenizer.from_pretrained("tomvaillant/qwen3-4b-journalist-ONNX");
const model = await AutoModelForCausalLM.from_pretrained("tomvaillant/qwen3-4b-journalist-ONNX", {
  dtype: "q4f16",
  device: "webgpu",
});

const messages = [
  { role: "user", content: "How do I find who registered a domain?" },
];

const inputs = tokenizer.apply_chat_template(messages, {
  add_generation_prompt: true,
  return_dict: true,
  enable_thinking: false,
});

const output = await model.generate({ ...inputs, max_new_tokens: 512 });
console.log(tokenizer.decode(output[0], { skip_special_tokens: true }));

Training

• Method: QLoRA (4-bit NF4 + LoRA r=16, alpha=32) via Oumi
• Data: 700 instruction/response pairs across 10 categories -- see training dataset
• Epochs: 1
• Hardware: A10G (HF Jobs)

ONNX conversion

• Quantization: q4f16 (4-bit weights, float16 activations)
• Runtime: Transformers.js via WebGPU
• Model size: ~2.5 GB quantized (from ~8 GB fp16)

Available files

• onnx/model_q4f16.onnx -- quantized model weights
• onnx/model_q4f16.onnx_data -- external data file
• Tokenizer files (BPE): tokenizer.json, merges.txt, vocab.json

Sources

Same training corpus as the Gemma variants. Covers OSINT tool selection, verification methodology, financial investigation (Follow the Money), digital security, media ethics, and investigative storytelling.

Key sources include the OSINT Navigator Tool Database (7,500+ tools), Bellingcat guides, GIJN manuals, UNESCO/Al Jazeera/CiFAR handbooks, SPJ ethics, RCFP legal resources, and Buried Signals investigation skill repositories.

Full attribution: SOURCES.md