Skip to content
Other

Tiny Aya Global Steering Vectors

inference-time LLM refusal control

2 min readupdated

A set of research steering vectors for CohereLabs/tiny-aya-global from sidg767, extracted using SafeTune. Unlike standard checkpoints or LoRAs, these files leave the model weights untouched—instead, they intervene directly in the residual stream during generation to make the model refuse unsafe prompts more often. The archive includes a static vector for layer 18 and a set of conditional vectors for CAST (Conditional Activation Steering, which triggers only when a specified condition is met). This release is geared toward AI safety researchers and developers exploring interpretability and LLM behavior control without retraining.

Key features

  • Not a standalone model or fine-tune: the base CohereLabs/tiny-aya-global weights remain completely frozen.
  • Supports two modes: a static shift at layer 18 (with α=4) and conditional CAST steering gated at layer 24.
  • Includes a ready-to-use set of 6 conditional vectors for layers 8, 12, 16, 18, 20, and 24, along with layer24_candidate_config.json.

What you can do

  • Test static refusal steering on a local copy of CohereLabs/tiny-aya-global by tuning the intervention strength α at layer 18.
  • Compare the trade-off between safety and false positives (over-refusal) using the static vector versus conditional CAST steering.
  • Load the pre-extracted tensors from the .pt files into your own activation interpretability experiments with SafeTune.

Example

The model card doesn't include ready-to-run inference code, though it outlines how the artifacts connect via SafeTune: refusal_direction_layer18.pt applies a static activation shift at layer 18 (α=4), while cast_layer_scan_vectors.pt pairs with layer24_candidate_config.json for conditional steering after a threshold check at layer 24.

Read moreDetails
40-Toxic benchmark: the static vector dropped unsafe responses from 2/40 to 0/40, but pushed Hard-set refusals up from 19/40 to 26/40; CAST kept the metric at 2/40.
Limitations: scores rely on a classifier rather than universal safety guarantees, and exact model and tokenizer revisions aren't pinned.
License: no license is declared for the artifacts themselves (SafeTune uses LSAL v1.2), while the base model is governed by Cohere 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.

    # sidg767/when-to-steer-tiny-aya
    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 sidg767/when-to-steer-tiny-aya --local-dir ~/models/when-to-steer-tiny-aya
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to when-to-steer-tiny-aya
  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. when-to-steer-tiny-aya

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

sidg767/when-to-steer-tiny-aya
Author
sidg767
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

BFL

BFL