你信任的每个基准可能已在训练数据中

Member-only story Artificial Intelligence Machine Learning LLM NLP Towards Ai Every Benchmark You Trust Is Probably in the Training Data by Now OpenAI admitted GSM-8K’s training set went into GPT’s training data. A researcher built a benchmark from an exam released two hours earlier specifically to avoid this. Aqeel Abbas 8 min read · 1 day ago -- Listen Share Press enter or click to view image in full size Photo by BoliviaInteligente on Unsplash In its own GPT-4 technical report, OpenAI disclosed something worth reading twice. They’d included portions of the training sets for MATH and GSM-8K — two of the most widely cited math reasoning benchmarks in the field — directly into GPT-4’s training data, specifically to improve performance on mathematical reasoning [1]. Read that plainly: the company reporting the benchmark score also reported, in the same document, that the score can’t be interpreted as zero-shot performance, because part of the answer key was in the study materials. They also disclosed a separate finding — that portions of the BIG-bench benchmark had been inadvertently mixed into the training set, contaminated enough that they excluded it from evaluation entirely rather than report a compromised number [1]. Two different failure modes, disclosed by the same company, in the same report: deliberate inclusion for capability gain, and accidental inclusion through the ordinary mechanics of training on a large chunk…