Determining LLM training-data exposure to evaluation datasets

Ascertain whether GPT-3.5-turbo, PaLM2 (chat-bison-001), and Falcon7b-instruct were exposed during pretraining, fine-tuning, or reinforcement learning from human feedback to the HOT toxicity dataset, SST-5 sentiment dataset, RumorEval dataset, and GVFC news frame dataset used in this study, to enable rigorous evaluation without data leakage.

Background

Rigorous evaluation of LLMs is complicated by the possibility that their training data may include commonly used benchmark datasets. If exposure occurred, performance estimates could be inflated or confounded, undermining the validity of empirical comparisons and conclusions.

The authors explicitly state that they cannot determine whether the models they evaluated were trained on or otherwise exposed to the datasets used in the experiments, highlighting a critical uncertainty that affects reproducibility and evaluation integrity.

References

Conducting rigorous evaluation of LLMs is challenging because we cannot determine whether these models have been exposed to our chosen datasets during their training phases, particularly popular datasets like SST-5.

— Prompt Design Matters for Computational Social Science Tasks but in Unpredictable Ways  (2406.11980 - Atreja et al., 2024) in Section 6 (Limitations)

As the exact training subsets are not publicly disclosed, possible overlap with IEMOCAP or MELD cannot be determined.

— Multimodal Emotion Recognition in Conversations via Class-Wise Adaptive Modality Fusion and Affective Geometry  (2609.09924 - Marín et al., 9 Sep 2026) in Section 3, Method, subsection “Feature Extraction and Modality Encoding,” paragraph “Audio features”

Because the complete pretraining corpora are not publicly available, it is not possible to exclude with certainty that the evaluated models were exposed to OhioT1DM data, derived material, or closely related distributions during pretraining . Such contamination cannot be quantified within the present study.

— It's All in the Way You Say It: The Role of Information Representation in LLM-Based Glycemic-Event Prediction  (2609.08772 - Apicella et al., 8 Sep 2026) in Section 6, “Limitations and future work”

Lastly, a general limitation of our study is that we cannot determine whether the evaluated LLMs were exposed to our benchmark datasets during pre-training.

— Does task decomposition improve automatic NLG evaluation?  (2609.01139 - Steindl et al., 1 Sep 2026) in Section Limitations