Safety Doesn't Translate

Safety Doesn't Translate

Shows side by side how Cohere's Tiny Aya answers before and after a safety repair that needs no retraining, using pre-judged replies from a test set. It suits AI safety researchers and multilingual model builders.

Made by DhruvParmar051 on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This is a static web page that compares two versions of Cohere's Tiny Aya model. The left side shows the model after a fine-tune that mixed in 100 harmful English examples, which lowered its safety in other languages too. The right side shows the same model after a repair done directly on its weights, with no retraining.

Who it suits: It suits people who study AI safety across languages, and builders who tune multilingual models and worry that a small change can hurt safety everywhere.

Industry | Who uses it | How they use it | Example
AI safety | Safety researchers | Compare drifted and repaired replies from the same test set | A researcher reads both sides for one prompt and notes the difference.
Model development | Teams tuning multilingual models | See how a small harmful fine-tune spreads to other languages | A team studies the drift chart before planning a fine-tune.
Policy and review | Reviewers and auditors | Read pre-judged evidence instead of running a model | A reviewer checks the figures and the excerpts.
Education | Students | Learn what safety drift is through concrete replies | A student opens the page and reads one example pair.

If you want to try your own prompts, skip it, because the page takes no live prompts.

Where to use it:

• A web browser, through the Space's page on Hugging Face
• Any browser on Windows, macOS or Linux, opening a downloaded copy of index.html

Use case | Who | What they get back
Compare two models | Researchers | Drifted and repaired replies side by side for test examples
See the charts | Anyone | A drift heatmap, a trade-off chart and an image of a drifted reply
Find the full results | Reviewers | A pointer to the project repository and the article

How to set it up:

• A web browser
• Python and pip, only for the download step

1. Install the huggingface_hub package.
2. Download the Space into a local folder named safety-doesnt-translate.
3. Move into that folder.
4. Open index.html in your web browser.

pip install -U huggingface_hub
hf download DhruvParmar051/safety-doesnt-translate --repo-type space --local-dir safety-doesnt-translate
cd safety-doesnt-translate

How to use it: Open the page and read the left side, labeled drifted, against the right side, labeled repaired. Look at the charts in the same page. The page has no input box.

Tool | What it does
index.html | The comparison page itself
drift_heatmap.png | A chart image of how the safety drift spreads
tradeoff.png | A chart image of the trade-off
via_english_drifted-harmful.png | An image tied to a harmful drifted reply

Example: You open the page and read the same test example on both sides. The left reply comes from the drifted model and the right one from the repaired model.

For newcomers and experts: If you are new, open the page and read one example on both sides. Drift means the model got worse at refusing harmful requests after a small change.

If you know the field, the models named are Tiny Aya Global and Aya Expanse 8B. Every reply was judged in advance. Flagged replies, and all replies in Swahili, Bengali or Filipino, are cut to short excerpts, because the judge does not cover those languages well.

Fine-tune with 100 harmful English examples
The model drifts in every language
A weight-space repair is applied
Pre-judged replies are shown side by side

Good to know:

• Replies are cut to excerpts when any judge flagged them
• Replies in Swahili, Bengali and Filipino are always cut, as the judge may miss harm there
• The full code, data and results are in the project repository and article, not in this Space
• License: cc-by-nc-4, where NC stands for non-commercial
• Cost to run: none named
• Data it reaches: only the pre-judged replies and images in its own folder
• Not for: live prompts, which the page leaves out on purpose

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: cc-by-nc-4.0

Links

• Space: https://huggingface.co/spaces/DhruvParmar051/safety-doesnt-translate
• Files: https://huggingface.co/spaces/DhruvParmar051/safety-doesnt-translate/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.