Exp_040 on Local First AIhttps://localfirstai.eu/tags/exp_040/Recent content in Exp_040 on Local First AILocal First AIhttps://localfirstai.eu/images/og-default.pnghttps://localfirstai.eu/images/og-default.pngHugoen-usSat, 17 Oct 2026 00:00:00 +0000Our 4B Property Model Gained 6.7 Points at Q4 and None at Q8https://localfirstai.eu/posts/2026-10-17-checkable-reward-gain-vanished-at-q8/Sat, 17 Oct 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-10-17-checkable-reward-gain-vanished-at-q8/We retrained Qwen3-4B on Spanish property law with GRPO against a reward that checks the answer itself: the final line must match a closed label. Our audit found the reward not gamed. At our Q4 the trained model gained 6.7 points of exact match on 150 fresh sealed questions; at Q8 the gain was 1.3 points and not significant. The same weights move about as many answers between Q4 and Q8 as training did.