Incorrect Context Uncertainty

To evaluate the robustness and context sensitivity of large language models (LLMs), we construct a dataset by augmenting the widely-used HotpotQA and Natural Questions (NQ) datasets with incorrect context passages. For each original question in these datasets, we generate a corresponding variant where the retrieved or provided context is intentionally misleading, irrelevant, or factually incorrect with respect to the question. This allows us to analyze the influence of incorrect contextual grounding on model behavior. Each question in the dataset includes: The original question (from HotpotQA or NQ), An incorrect context passage (generated by GPT4.1 mini), The gold answer from the original dataset (used only for reference), The generated responses + logits + probabilities + additional metadata from three LLMs: Fanar-9B: Our in-house Arabic-English bilingual model, Qwen-7B-Chat: A state-of-the-art open-source LLM from Alibaba, Gemma-13B-IT: A strong instruction-tuned model from Google.