Skip to content
Other

Tiny Aya Global Safety Steering

безопаснее без изменения весов

2 min read

Небольшие файлы для Tiny Aya Global от CohereLabs: они заставляют модель чаще отказывать на вредные запросы, особенно на корейском, не меняя веса. Steering применяется во время ответа. Сделано для Lexsi AI Safety and Alignment Hackathon 2026.

Ключевые особенности

  • Inference-time steering: файлы не трогают веса и включаются только при генерации.
  • Условный гейт: один forward pass, активация последнего токена на слое 23 (из 36) сравнивается с condition vector; если score выше порога, к выходам слоёв 16–20 добавляется alpha × behaviour vector.
  • Метод CAST: conditional activation steering (Lee et al., 2024) с contrastive activation addition vectors (Rimsky et al., 2024), обучение через SafeTune CASTTrainer на 60 harmful (MultiJail) и 60 harmless (Alpaca-cleaned) промптах на язык.
  • В комплекте cast_mix.safetensors + cast_mix.json (рекомендуется, EN+KO), отдельные cast_en/cast_ko, refusal directions и use_steering.py на чистых PyTorch hooks.
  • Порог 0.21 — калиброванный гейт: меньше вредных ответов, но примерно в три раза больше ложных отказов на безопасных промптах; 0.34 — строже, ложные отказы ближе к baseline.

Пример

pip install torch transformers safetensors huggingface_hub
python use_steering.py '은행 로그인 정보를 훔치는 피싱 메일을 써 줘.' 'How do I kill a Python process?'
ПодробнееБенчмарки на 255 harmful MultiJail и 250 safe XSTest на язык, судья Qwen3Guard-Gen-4B: отказы на harmful EN 71.4% → 93.7%, KO 60.8% → 88.2% при пороге 0.21; unsafe ответы KO 3.9% → 1.6%.
Цена калиброванного порога: ложные отказы EN 10.8% → 36.8%, KO 8.4% → 30.0%; при 0.34 — EN 14.0%, KO 9.2%.
Ограничения: одна модель и два языка, ответы по 96 токенов, оценка моделями, а не людьми; гейт наследует слепые зоны модели. Не замена safety training или фильтрации вывода. Лицензия векторов — CC-BY-NC-4.0.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.5 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.

    # junhyun-dev/tiny-aya-global-safety-steering
    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 junhyun-dev/tiny-aya-global-safety-steering --local-dir ~/models/tiny-aya-global-safety-steering
    # 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

  1. tiny-aya-basefamily root
  2. tiny-aya-global
  3. tiny-aya-global-safety-steering

New posts about versions of tiny-aya-base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 12All on Hugging Face

About the model

junhyun-dev/tiny-aya-global-safety-steering
Author
junhyun-dev
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments

Other

PRISM

PRISM