https://localfirstai.eu/tags/lora/Recent content in Lora on Local First AIhttps://localfirstai.eu/images/og-default.pnghttps://localfirstai.eu/images/og-default.pngHugoen-usThu, 15 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-15-code-checked-reward-did-not-make-it-more-correct/Thu, 15 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-15-code-checked-reward-did-not-make-it-more-correct/We retrained Qwen3-4B on Spanish property law with GRPO against a reward computed only by code: no teacher answers, no judge in the loop. The first run gamed the reward and our own audit stopped it. The second passed the audit and seemed to gain +0.20 in correctness, but a control run at less compression showed the gain came from our quantiser, not from training. The model learned to cite what it retrieved. It did not learn the answer format, and it did not get more correct.