Pert Lomp

Low-resource language post-training on consumer hardware. I teach open LLMs proper Estonian on a single RTX 5090: locked evals, rule-based data generation validated with morphological analysers (Vabamorf/EstNLTK), surgical SFT rounds built from the model's own errors, on-policy DPO, and continued pretraining on edited prose. Everything measured, every failure documented, all methods and adapters released openly. Mission: a small language should be as good in open models as in paid ones — and this work should carry forward into future models. Non-commercial.