Skip to content
Computer vision

WSP safety edit Qwen3-VL-4B

отказ от мультимодальных джейлбрейков

2 min readupdated

Исследовательский чекпоинт из работы по weight-space programming (WSP): LLM-агент напрямую правил веса Qwen3-VL-4B-Instruct, чтобы модель чаще отказывалась от мультимодальных джейлбрейк-запросов. Обучения, градиентов и оптимизатора не было; модель осталась в BF16 с теми же именами тензоров и шардов, что у базы. Размер — 4,44 млрд параметров.

Ключевые особенности

  • В MLP слоёв 20 и 26 добавлены два маленьких «refusal circuit»; каждый задействует два наименее активных нейрона слоя.
  • Порог срабатывания задан почти константным residual-направлением, а выход пишет собственное refusal-направление модели с противоположными знаками.
  • На VLGuard test ASR для всех harmful-запросов упал с 34,5% до 16,5%; для safe image + unsafe instruction — с 22,6% до 5,0%.
  • Ни один harmful-промпт, от которого база отказывалась, не стал отвеченным; over-refusal на безопасных запросах вырос с 0,36% до 1,79%.
  • Загружается как обычная Qwen3VLForConditionalGeneration без патчинга; изменено около 30,7 тыс. из 4,44 млрд весов.

Пример

import torch
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
repo = 'utsavz/wsp_nov10_qwen3vl4b_opus'
model = Qwen3VLForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16, device_map='auto')
processor = AutoProcessor.from_pretrained(repo)
messages = [{'role': 'user', 'content': [{'type': 'image', 'image': 'photo.jpg'}, {'type': 'text', 'text': 'Describe this image.'}]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors='pt').to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
ПодробнееVLGuard test, 1 558 samples: ASR all harmful 34,5% → 16,5%, over-refusal 0,36% → 1,79%.
Изменено ~30,7 тыс. из 4,44 млрд весов; остальные 707 тензоров, конфиги и токенизатор бит-в-бит совпадают с базой.
Ограничения: keyword-скоринг ASR, не покрыта обфускация (leetspeak), возможен лексический over-refusal, benign-ответы меняются; только research use, не production safety system.
Discuss Telegram 0 11

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.3 GB · ready files9.6 GB
  • FP8 / INT8recommendedweights 4.1 GB5.0 GB
  • 4-bitweights 2.3 GB3.1 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 3080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # utsavz/wsp_nov10_qwen3vl4b_opus
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download utsavz/wsp_nov10_qwen3vl4b_opus --local-dir ~/models/wsp_nov10_qwen3vl4b_opus
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-VL-4B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 5All on Hugging Face

About the model

utsavz/wsp_nov10_qwen3vl4b_opus
Author
utsavz
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

110

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Language models

Niti Gate

Niti Gate