GPT-NL: building the annotation platform behind the Dutch national language model

GPT-NL is the Dutch national language model, developed by TNO, NFI, and SURF. For the finetuning phase, thousands of high-quality Dutch instruction prompts had to be created by hand. No existing dataset met the strict requirements for clean, legally obtained data. The AI Factory built the annotation platform that made this possible, and Spectrum Intelligence (SPIN.AI) used it to coordinate a team of annotators who crafted every single prompt.

The challenge: finetuning data from scratch

Finetuning a language model requires thousands of carefully crafted instruction-response pairs. For GPT-NL, no suitable Dutch prompt dataset existed that guaranteed a fully clean data chain. Every prompt had to be written by hand to ensure no AI-generated or unlicensed content contaminated the training data.

The target: approximately 15,000 prompts across eight categories including open questions, closed questions, chat, creative writing, classification, brainstorming, and summarization. Each prompt needed a corresponding high-quality completion. This required a team of 10 to 15 annotators working in structured, agile sprints with continuous feedback loops.

The annotation platform we built

The AI Factory designed and built the annotation software that powered this entire process. The platform was engineered specifically for the demands of LLM finetuning: managing large annotation teams, enforcing quality standards, and maintaining clean data provenance at every step.

The platform provided:

Structured workflows for creating instruction-completion pairs, with built-in category management across all eight prompt types.

Quality assurance pipelines with automated validation, duplicate detection, and cross-review capabilities between annotators.

Real-time progress tracking and analytics dashboards, giving project managers full visibility into annotation throughput and quality metrics.

Complete data provenance logging, ensuring every prompt can be traced back to its human author. Critical for GPT-NL's clean data chain commitment.

Collaborating with Spectrum Intelligence

Spectrum Intelligence (SPIN.AI) brought the annotation team: 10 to 15 annotators, primarily people on the autism spectrum. What makes this collaboration special is that people with autism often have exceptional skills in precision, focus, and attention to detail. Exactly what high-quality data annotation demands.

SPIN.AI combines a strong social mission with high-quality AI work. CEO Michael Radvany founded the company to offer meaningful employment to one of the most underestimated populations in the workforce. Approximately 80-90% of people on the autism spectrum in the EU are unemployed. SPIN.AI proves that their talents are not just valuable, but essential for work that requires the precision most people cannot sustain.

Quality through structured feedback

High-quality finetuning data does not happen in one pass. Our platform was built around an iterative feedback loop: annotators create instruction-completion pairs, reviewers evaluate them against quality criteria, and rejected items cycle back with specific guidance on what to improve. This tight loop is what turns raw annotations into training data you can trust.

The agile workflow between GPT-NL, Spectrum Intelligence, and The AI Factory's platform enabled rapid iteration. Project managers could track quality metrics in real-time, identify patterns in rejection reasons, and refine annotation instructions on the fly. The result: a dataset that improved with every sprint, not just in size but in consistency and depth.

A clean data chain

GPT-NL is committed to training exclusively on legally obtained data. By creating all prompts by hand through our annotation platform, the project guarantees that no AI models trained on unlicensed data were used in the finetuning process. Every prompt is traceable to its human creator.

This is not just an ethical choice. It is a requirement for building a national language model that organizations and government institutions can trust and deploy with confidence.