Skip to content
Computer vision

Qwen3-VL-2B + WSP

правка безопасности без обучения

2 min readupdated

Чекпоинт принимает изображение и текстовый запрос и отвечает текстом, как обычная vision-language модель. Он основан на Qwen3-VL-2B-Instruct (2,13 млрд параметров, BF16) и отличается точечной правкой весов: без градиентов, файнтюна и оптимизатора изменены 7 тензоров. Цель — снизить вероятность вредных ответов на опасные изображения и инструкции, сохранив нормальные ответы на безопасные запросы. Пока это research checkpoint: независимой проверки на held-out-наборе нет, а результаты ниже получены только на dev-сплитах VLGuard. Пригодится исследователям safety для VLM, а не как готовый продакшн-фильтр.

Ключевые особенности

  • База — Qwen3-VL-2B-Instruct; сохранены layout, tokenizer, processor, chat template и BF16, изменены только 7 тензоров.
  • Метод complementary WSP: якорь — gated harm-detector нейрон в layer 15, плюс rank-1 правка attention в layer 18 и gated image-borne-harm нейрон в layer 10.
  • На confirmation-сплите VLGuard train ASR harmful снизился с 37,5% до 10,0%, unsafe-image — с 50% до 20%, unsafe-instruction — с 25% до 0%.
  • Over-refusal остался 0/24; benign reference ΔNLL +0,010. Но это малые dev-выборки и keyword-based ASR.

Что можно сделать

  • Прогнать собственные запросы с изображениями и текстом и сравнить отказы с оригинальной Qwen3-VL-2B-Instruct.
  • Изучить, как правка отдельных нейронов и attention-слоёв влияет на safety-поведение VLM.
  • Использовать как research-чекпоинт в экспериментах с guardrails, не как готовый продакшн-фильтр.

Пример

import torch
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
model = Qwen3VLForConditionalGeneration.from_pretrained("utsavz/wsp_nov11_qwen3vl2b_complementary", dtype=torch.bfloat16, device_map="cuda")
processor = AutoProcessor.from_pretrained("utsavz/wsp_nov11_qwen3vl2b_complementary")
messages = [{"role": "user", "content": [{"type": "image", "image": "img.jpg"}, {"type": "text", "text": "Describe this image."}]}]
inputs = processor.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])
ПодробнееНа confirmation-сплите VLGuard train (64 примера на подмножество): ASR harmful 37,5% → 10,0%, unsafe-image 50% → 20%, unsafe-instruction 25% → 0%; over-refusal 0/24.
Benign reference ΔNLL +0,010. Ограничения: нет held-out оценки; вредные изображения при нейтральном запросе всё ещё не отклоняются (0/32); маленькие выборки, keyword-based ASR и лимит 20 токенов.
Правка не заменяет safety-обучение и внешние guardrails.
Discuss Telegram 0 15

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 4.0 GB · ready files4.9 GB
  • FP8 / INT8recommendedweights 2.0 GB2.7 GB
  • 4-bitweights 1.1 GB1.7 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 2060

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_nov11_qwen3vl2b_complementary
    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_nov11_qwen3vl2b_complementary --local-dir ~/models/wsp_nov11_qwen3vl2b_complementary
    # 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-2B-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 2All on Hugging Face
Merges 1All on Hugging Face

About the model

utsavz/wsp_nov11_qwen3vl2b_complementary
Author
utsavz
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

150

Ask AI about this model

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

Comments