CourseLarge Language Models: The Complete Course · Module 9: Fine-Tuning and Adaptation · part 67 of 103
Part 67 · Module 9: Fine-Tuning and Adaptation

Incident: Reworded test questions passed the leak check

17 min read·13 Oct 2026
Module 9 · Incident file · About 30 minutes

What happened

NOVEMBER 2023: ONE PAPER, ONE BLOG POST 8 Nov arXiv v1 11 Nov v2 revised 14 Nov LMSYS blog reworded test questions pass n-gram checks
Three dated publications (blue) carry one finding (red): string checks let reworded test questions into training.

In November 2023 Shuo Yang, Wei-Lin Chiang, Lianmin Zheng, Joseph E. Gonzalez and Ion Stoica (LMSYS, UC Berkeley) published "Rethinking Benchmark and Contamination for Language Models with Rephrased Samples" (arXiv 2311.04850, 8 November, revised 11 November), with a blog post on 14 November titled "Catch me if you can! How to beat GPT-4 with a 13B model".

Their target was decontamination: the check that removes copies of test questions from training data. The usual checks compare strings, such as shared 10-word runs. The authors reworded or translated benchmark test questions and fine-tuned Llama-2 and CodeLlama models on them. The paper reports that "both rephrased and translated samples are undetected by the n-gram overlap".

TRAINED ON REWORDED TEST QUESTIONS (PAPER'S NUMBERS) GSM-8K, Llama-2 13B 28.7 before 95.3 HumanEval, CodeLlama 13B 36.0 before 81.1 GPT-4 on HumanEval: 67.0
Scores before (grey) and after training on reworded test questions (red): a 13B model passed GPT-4 by having seen the questions.

The scores rose sharply. On GSM-8K, Llama-2 7B and 13B went "from 28.7 to 95.3". On HumanEval, CodeLlama 13B went from 36.0 to 81.1, while "GPT-4 can only achieve 67.0". The blog reports MMLU 85.9 for a 13B model. None of this was new ability: the models had seen the questions.

SHARE OF HUMANEVAL FOUND IN TRAINING SETS The Stack (subset) 18.9% StarCoder-Data (subset) 15.9% CodeAlpaca 12.8% RedPajama-1T (subset) 8.5%
Reworded test questions (red) were already inside widely used training data, measured on the subsets the paper scanned.

The cost was in public data that others train on. Their LLM-based checker found test questions, reworded, in widely used training sets: in CodeAlpaca, "21 rephrased samples from the HumanEval test set, accounting for 12.8%", and, in the abstract, "8-18% of the HumanEval benchmark" overlapping RedPajama-Data-1T and StarCoder-Data. Their advice: use stronger decontamination, and "develop fresh one-time exams".

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Buy this course · ₹399See what is inside