GlobalAnnotate logoGlobalAnnotate
All case studies
AI AnnotationAI / LLM10-week programme

Multilingual RLHF dataset for an enterprise LLM

A Series-B foundation-model company

An 80k-prompt preference dataset across 8 languages — including red-team adversarial prompts and instruction-following evals — that lifted the model's downstream evals by 11.6%.

80k
Prompts
8
Languages
+11.6%
Eval lift

This case study is part of our AI Annotation work — see how the same approach scales for other teams in our portfolio.

Challenge

What we walked into.

The model team needed preference data that captured real instruction-following nuance across eight languages, including coverage of adversarial and safety-critical prompts. Existing public datasets were thin outside English and inconsistent in quality.

What we did

The work, step by step.

  1. Designed the dataset schema with the research team — preference pairs, instruction-following labels, and adversarial prompts

  2. Recruited native annotators with subject-matter expertise in every target language and ran calibration on a held-out gold set

  3. Generated 80k prompts across 8 languages with a balanced mix of intent categories and difficulty bands

  4. Shipped per-language IAA reports and adjudicated edge cases with the research team in weekly review sessions

Results

What it shipped.

Outcomes measured against the brief we agreed up front, not vanity metrics.

  • Prompts
    80k
  • Languages
    8
  • Eval lift
    +11.6%

准备好拓展全球市场了吗?

告诉我们您的项目需求,我们将在一个工作日内回复。