Reader's Comment Exposed a Hidden Flaw in an AI Benchmark Comparison

A developer published results from a LoRA fine-tuning experiment comparing fine-tuned and prompted language models, initially concluding that fine-tuning offered little advantage. A reader named Max Quimby commented that the unequal performance drops between methods suggested the original test set was easier for the prompted model, prompting the author to investigate further. The author discovered a more fundamental flaw: the fine-tune row in the comparison table actually represented two different adapters trained on different datasets, making the five-point drop misleading. Running the original v1-trained adapter against the newer real-world test set revealed a 33-point drop, larger than any prompting method. The corrected data showed a clear pattern — the more a method had been fitted to the original data distribution, the more its accuracy fell when tested on real-world data.