With an MIT feasibility grant via RVO, we investigated whether a reinforcement-learning pipeline can improve compact, open-source language models in a targeted way for finding and answering questions within a Dutch, domain-specific context. The study ran from May 2025 to January 2026 and combined technical development with a market exploration.
We built a full, reproducible training pipeline: splitting a document into small, traceable text blocks, having context written per block, fine-tuning a compact model on that (supervised, with LoRA) and then steering it with PPO on the judgement of a separately trained evaluation model. The pipeline is modular, checkpointed and restartable, so the dataset and models are interchangeable per domain without rebuilding the chain.
The evaluation on public benchmarks confirms the core mechanism: better context demonstrably leads to better retrieval, with a stepwise improvement in recall and precision. PPO optimisation on AI preferences, however, did not on its own deliver better results than supervised fine-tuning alone. Assessing context quality turned out to be a difficult task for an AI judge, and a follow-up would use retrieval quality directly as the reward signal.
Alongside the technology, we carried out a market exploration with fourteen (semi-)public organisations. That produced a consistent picture: broad demand for compact, task-specific AI on Dutch, open-source and auditable infrastructure, with sovereignty as a precondition and willingness to pay tied above all to demonstrable, measurable value.
The conclusion is positive and nuanced: the core of the approach is feasible, and the targeted use of reinforcement learning on narrow, reusable components is technically and commercially promising, provided that measurability and validation are at its heart.
