Case studies
Turkish language data and model alignment for a global AI developer
For a global AI developer whose models underperformed on Turkish, a workforce hired for morphological judgement and an error taxonomy moved the evaluation.
The challenge
The client's models handled Turkish measurably worse, and the evaluation set showed why: morphology. English-tuned tokenisation shredded meaningful units and generated suffix agreement errors no native speaker would make. Previous vendors produced data that looked acceptable to non-speakers and failed on evaluation.
What Corpshore did
We recruited a distributed Turkish workforce with an unusually high linguistic bar. The programme ran in four phases. The substantive contribution was the error taxonomy: annotators categorised failures by linguistic cause, giving the research team a signal to act on. Quality was managed as a gate, not a report.
Results
- Delivered several hundred hours of transcribed Turkish speech plus large volumes of preference and evaluation data, all passing acceptance
- Inter-annotator agreement held above threshold through two scale increases
- The client reported measurable improvement on their internal evaluation set, largest on morphological accuracy
- The programme extended twice and expanded into Azerbaijani and Uzbek
Why it worked
The client had been buying volume. What it needed was linguistic diagnosis. Recruiting for morphological judgement cost more per hour and produced data that actually moved the evaluation.
Let us build your Türkiye team
Tell us your function, your scale and your language needs. We will come back to you within six hours.