Overestimation in LLM Evaluation: A Controlled Large-Scale Study on Data Contamination’s Impact on Machine Translation

We thank the reviewer for their positive review. We will extend the discussion section to include a discussion on the broader implications of our findings.

On the choice of BLEU as a metric: We acknowledge that a string-based metric (like BLEU) has limitations. For this reason, we already accompany all of our evaluations with an additional learned metric (i.e., MetricX) which has been widely adopted by the field. All drawn conclusions are consistent across both reported metrics. Our choice to surface BLEU in the main paper is based on its wider adoption within the ML community.

Thanks a lot for your review of our paper.

First, we respectfully disagree that the contribution of our paper is insignificant. The key assumption of data contamination is motivated by the hypothesis that consuming test data leads to an overestimation of a model’s performance. While intuitive, this hypothesis has not been rigorously tested at this scale, and our work aims to fill that gap in the literature. Further, we suspect many researchers care about testing this hypothesis but are unable to run the experiments that we did due to the computational cost. Therefore, we feel our contributions are significant and our findings are very relevant to the research community.

1. There are methods for detecting contamination by looking at the overlap between the training and test sets, including https://arxiv.org/abs/2411.03923. Our train-test decontamination stage is inspired by the findings of this paper. Our implementation is very efficient because it’s highly parallelizable and uses a Bloom filter combined with an exact search to search more than 100k 8-grams from the test data in trillions of training tokens in less than a day.

2. There are papers focused on finding contamination with just access to model activations. However this task is shown to be harder than initially expected (https://arxiv.org/abs/2402.07841.) Our experiments and findings regarding contamination in section 5.4 can partially highlight why this task can be harder than initially hypothesized. We show that models don’t naively memorize text in the pretraining data suggesting that model behaviour might not be as easily separable between what is in the pre-training data and what is not. Recent work also shows that models can verbatim generate text that is explicitly not in their pre-training data https://arxiv.org/abs/2503.17514. This suggests that models learn more general representations of language and understanding contamination and the impact on evals requires understanding knowledge formation and access in large language models.

We thank the reviewer for the detailed review and insightful comments. We will use the reviewer’s feedback to revise the writing of Section 3.4 and include missing citations.

1. Details on contamination score calculation: The contamination score is calculated as the number of tokens in the longest overlap, over to the number of tokens in the eval example. If more than 70% of the tokens are contained in the longest matching overlap we label an example as contaminated and remove it from our test set.
2. Choice of WMT23/24 datasets: The WMT23/24 datasets used in our experiments are the standardized datasets, established by WMT (Conference on Machine Translation) annual conference.
3. Contamination from WMT’24: Source segments from WMT’24 are generally standard pieces of texts that are frequently found in news articles or general webtext. Since WMT sources input segments from online sources, it is possible that frequently used sentences might be found in any pre-training data. For this reason, we make sure that we run decontamination tests for all datasets we use.
4. On the ranking of Figure 2: The presented methods are sorted based on the average performance across language pairs. They are shown in descending order for both Figs 2 and 3. We will clarify this in the next version of the paper.
5. Checkpoint branching details: In our branching method, each branch is effectively a copy of the baseline model, with all the optimizer and meta data inherited from the checkpoint. As training continues, the only difference is the pre-training mixture that the new model (continued pre-training) is exposed to. This mixture is a copy of the baseline one with random examples replaced by the contaminated instances. Our setup covers 42 contamination experiments, leading us to 42 “branches”. We will include those details in the next version.

We thank the reviewer for the detailed comments and feedback. We respond to the reviewer’s comments below.

1. Question 1 (training compute): All models are implemented as continued pre-training starting off baseline checkpoints. For instance, models with late contamination are only trained for 10% of the entire training (given we kickstart the model from the 90% of the baseline’s mark). In practice, this means that our checkpoint branching approach reduces the compute requirements about 60% (compared to the compute budget we would have spent, if all models were trained from scratch. The experiments being expensive is definitely a challenge however we have undertaken this cost for the community and shared our findings openly that can answer critical questions in this area as well as inform future research. We think it is an important contribution of this paper that we have committed to rigorously experimenting the impact of contamination given the computational requirements and not a weakness.
2. Question 2 (generalization to different architectures): We acknowledge that our findings are limited to decoder-only LLMs, which constitute the predominant LLM paradigm in recent years.
3. Question 3 (frequency of contamination in pre-training): While in principle, pre-training data deduplication is a common practice; in multilingual settings, upsampling under-represented languages is an equally common tactic. The latter is dictated by the lack of resources for many languages, and inevitably can lead to data being repeated more than once. As a result, it is reasonable to assume that once a test instance bleeds into pre-training data, it might end up seen by the model more than once.
4. On the novelty of our findings: The more intuitive result, i.e., contamination leads to performance overestimation, has been hypothesized by many works; however, which data conditions contribute to this overestimation and to what extent, has not been studied by prior work. Our study contributes empirical evidence that contamination could, but not always lead to performance overestimation, a behavior that depends on various factors ranging from the way contaminated test sets are presented in the pre-training data, to how well the contaminated languages is represented in the pre-training distribution (Section 5 and 5.1).
5. On the choice of Machine Translation (MT) test sets: MT test sets naturally let us study how data contamination interacts with pre-training resource requirements (which are different across the many languages we studied with this work). As evidenced from our findings, whether contamination has a measurable impact on downstream performance does connect to the language’s representation during pre-training.