RLHF, adversarial prompts and toxicity annotation for safe GenAI

Making a generative model safe in one language is hard. Making it safe across many is a data problem — and a filter bolted on at the end won't do it. Our safety work combines three workflows that reinforce each other.

Adversarial prompt engineering

We author prompt–response data that includes adversarial prompts designed to surface failure modes — probing the model where it's weakest, not just where it's already strong.

Human-in-the-loop RLHF

Native-speaker preference and feedback signal drives RLHF for alignment. Human judgement, language by language, is what teaches the model what's actually acceptable to its users.

Toxicity annotation in native context

Toxicity detection and safety-focused annotation are grounded in native-speaker judgement, not translated taxonomies. What counts as harmful is culturally specific, and the annotation has to reflect that.

One workflow, not three silos

Adversarial data finds the failures, RLHF teaches the preferences, and safety annotation labels the harms — feeding scalable pipelines for fine-tuning and evaluation rather than living in separate stages.