qwen3.5-4b-privacy-defender

Qwen3.5-4B Privacy Defender

A privacy-preserving prompt-rewriting model. It rewrites a user's message so it can be sent to an external chatbot without revealing personal identity (location, profession, age, gender, family/relationship status, socioeconomic status), while preserving the original meaning and intent.

This is a fine-tuned derivative of Qwen/Qwen3.5-4B (SFT followed by DPO). It is not an official Qwen model.

Intended use

On-device rewriting of chat prompts before they leave the user's machine. Developed as the engineering artifact of a master's thesis on protecting user privacy in long-term AI conversations.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained("oberus/qwen3.5-4b-privacy-defender", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "oberus/qwen3.5-4b-privacy-defender", torch_dtype="auto",
    device_map="auto", trust_remote_code=True)

SYSTEM = ("You are a privacy-preserving rewriting assistant. Rewrite the user's message "
          "so it can be safely sent to an external AI chatbot without revealing the user's "
          "personal identity ...")  # full prompt in the thesis repo

messages = [{"role": "system", "content": SYSTEM},
            {"role": "user", "content": "Hi, I'm Ivan, a nurse in Boston."}]
enc = tok.apply_chat_template(messages, add_generation_prompt=True,
                             return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**enc, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))

License

Apache 2.0, inherited from the base model (Copyright 2026 Alibaba Cloud). This repository contains a modified (fine-tuned) version of Qwen3.5-4B.