Skip to content
Other

DeBiasLens

интерпретируемое устранение предвзятости в VLM

1 min readupdated

Набор чекпоинтов разреженных автоэнкодеров (SAE — моделей, раскладывающих внутренние активации нейросети на понятные признаки) от исследователей из KAIST AI. Они обучены на внутренних представлениях визуальных и визуально-языковых энкодеров для анализа и корректировки социальных смещений (bias steering). В основе лежат популярные базовые модели: clip-vit-base-patch16, clip-vit-large-patch14-336, InternViT-300M-448px и llava-onevision-qwen2-7b-ov-hf. Инструмент пригодится исследователям интерпретируемости и разработчикам, работающим над снижением предвзятости мультимодальных моделей.

Ключевые особенности

  • Архитектура Matryoshka BatchTopK SAE с параметром k=20, обученная на остаточных активациях после MLP-слоёв.
  • Поддержка четырёх популярных энкодеров: двух версий CLIP, InternViT-300M и LLaVA-OneVision-Qwen2-7B.
  • Разные коэффициенты расширения скрытого пространства (x1, x2, x4, x8) и обучение на датасетах fairface, celeba, bios, cocogender, cocogendertxt и sbbench-syn-crop.
  • В каждой папке лежат готовые веса ae.pt и файл config.json, а структура каталогов напрямую совместима с кодом обучения.

Что можно сделать

  • Проанализировать внутренние признаки и социальные смещения в конкретных слоях CLIP (11 и 23), InternViT (23) или LLaVA-OneVision (25).
  • Провести корректировку (steering) активаций VLM без переобучения всей базовой модели, опираясь на методику из статьи CVPR 2026.

Пример

from huggingface_hub import hf_hub_download
from dictionary_learning.trainers.matroyshka_batch_top_k import MatroyshkaBatchTopKSAE

path = hf_hub_download(
    "kaist-ai/debiaslens",
    "matroyshka_batch_top_k_20_x1/random_k_2/fairface_train_activations_clip-vit-large-patch14-336_23_post_mlp_residual_matroyshka_batch_top_k_20_x1/trainer_0/ae.pt",
)
sae = MatroyshkaBatchTopKSAE.from_pretrained(path)
ПодробнееПубликация: CVPR 2026 (arXiv:2602.24014).
Целевые слои: слой 11 (clip-vit-base-patch16), слой 23 (clip-vit-large-patch14-336, InternViT-300M-448px), слой 25 (llava-onevision-qwen2-7b-ov-hf).
Формат хранения: веса ae.pt и конфигурация config.json для библиотеки dictionary_learning.
Discuss Telegram 1

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 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.

    # kaist-ai/debiaslens
    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 kaist-ai/debiaslens --local-dir ~/models/debiaslens
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to debiaslens
  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 debiaslens (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

kaist-ai/debiaslens
Author
kaist-ai
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments

Other

HTR-500M

HTR-500M