AntiLeak-Bench: Preventing Data Contamination by Automatically Constructing Benchmarks with Updated Real-World Knowledge

🔼 The figure illustrates the automated process of building the AntiLeak-Bench. It starts with preparing data from Wikidata. The workflow then identifies knowledge updated after an LLM’s knowledge cutoff time by comparing claim histories. Next, supporting documents are retrieved from Wikipedia based on the updated knowledge. Finally, contamination-free question-answering samples are generated using the updated knowledge and supporting documents.

read the caption

Figure 2: Illustration of the automated benchmark building workflow without human labor. After data preparation, it includes three main steps: (1) Identify updated knowledge after the cutoff time; (2) Build supporting documents; (3) Construct contamination-free samples (Figure 3 exemplifies how to construct multi-hop samples).

🔼 The figure illustrates the process of constructing multi-hop question-answering samples. It starts with an initial fact, such as Lionel Messi being a member of Inter Miami. Subsequent ‘hops’ are made by connecting the object of the previous fact to the subject of a new fact, forming a chain. For example, the second hop connects Inter Miami to its location (or head coach), and a third hop might link the head coach to their country of citizenship. This chain of relations forms the basis of a multi-hop question, where the answer requires traversing multiple linked facts. The supporting context for the question would include text related to each entity involved in these ‘hops’.

read the caption

Figure 3: Illustration of constructing multi-hop samples. Find the consequent relation of previous objects.

🔼 This figure presents the Exact Match (EM) and F1 scores of different large language models (LLMs) over multiple 2-month or 3-month intervals between 2022 and 2024. The models evaluated include Llama-2-7B, Llama-2-13B, Mistral-7B, Vicuna-v1.5-7B, LongChat-v1.5-7B, Phi-3.5-mini, Qwen-2-7B, Mistral-Nemo-12B, and Gemma-2-9B. The x-axis represents the time intervals, while the y-axis represents the EM and F1 scores. Different colors and line styles distinguish the performance of each model. The vertical dotted lines likely represent the knowledge cut-off times of the LLMs, indicating the point in time after which information used in evaluating the models was not included in their training data. The figure demonstrates the performance trends of different LLMs over time, highlighting potential data contamination issues and illustrating the effectiveness of the AntiLeak-Bench in evaluating LLMs in a contamination-free environment. This figure is important for understanding the reliability and validity of using benchmarks to assess LLMs’ capabilities.

read the caption

Figure 4: EM and F1 performance at each time interval.

🔼 This figure showcases the proportion of times Large Language Models (LLMs) select correct and outdated options in multi-choice question answering tasks related to data contamination in evaluations. The figure is separated into two parts based on different models and time intervals reflecting knowledge cut-off dates and updates. The analysis reveals that outdated options were selected more frequently by LLMs over time, and LLMs struggled to answer the questions correctly, with some LLMs performing poorly even before their knowledge cut-off date. The x-axis represents different time intervals, while the y-axis shows the percentage. Each line in the chart represents the selection frequency of an option over each time interval.

read the caption

Figure 5: Correct and outdated option proportions at each time interval.

🔼 This table presents an example from the AntiLeak-Bench, demonstrating how questions, answers, and contexts are structured within the benchmark. It includes examples for both Generation and Multi-Choice question formats. The attributes provided are ‘question’ (in both formats), ‘answer’ (in both formats), ‘subject’, ‘pid’ (property ID), ‘object’, ‘object_old’, and ‘context’. The table showcases the different components used to create a contamination-free example.

read the caption

Table 2: An example from AntiLeak-Bench.

🔼 This table presents the human evaluation results of the generated samples’ context and answer accuracy for single-hop and multi-hop question answering. The results demonstrate high accuracy for both contexts and answers in the generated samples, indicating that the samples are of good quality.

read the caption

Table 3: Data quality by human verification.

🔼 This table presents the Exact Match (EM) and F1 scores for several Large Language Models (LLMs) evaluated on the AntiLeak-Bench using the generation format. The benchmark evaluates LLMs’ ability to answer questions about updated real-world knowledge, while mitigating data contamination. Results are reported for different conditions: ‘Gold’ signifies evaluations with only relevant supporting documents provided, while ‘Ndsubscript𝑁𝑑N_{d}italic_N start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT’ denotes evaluations with an increasing number (3, 5, or 7) of distracting documents included in the context. Higher EM and F1 scores signify better performance and the highest scores are highlighted in bold. This allows for an analysis of LLM performance under varying difficulty levels within the AntiLeak-Bench.

read the caption

Table 4: EM (Exact Match) and F1 results in the generation format on AntiLeak-Bench. Gold means only gold documents; Ndsubscript𝑁𝑑N_{d}italic_N start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT is the number of distracting documents. The best is in bold.

🔼 This table presents the accuracy (Acc) and F1 scores of several large language models (LLMs) on the AntiLeak-Bench using the multi-choice question format. The benchmark evaluates LLMs’ ability to answer questions correctly given a context, where ‘Gold’ refers to providing only the gold standard supporting document as context. ‘N_d’ represents the number of additional distracting documents added to the context, increasing the task’s difficulty by requiring the models to filter out irrelevant information. The table compares LLM performance across different levels of distraction (N_d = 3, 5, 7) and identifies the best-performing model for each setting with bold formatting.

read the caption

Table 5: Acc and F1 results in the multi-choice format on AntiLeak-Bench. Gold means only gold documents; Ndsubscript𝑁𝑑N_{d}italic_N start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT is the number of distracting documents. The best is in bold.

🔼 This table presents the number of samples within each time period, task, and number of distracting documents in AntiLeak-Bench. The table is split into two rows based on time period (2022-01-01 to 2023-01-01 and 2023-05-01 to 2024-08-01). The columns represent different tasks: single-hop and multi-hop question answering, with varying numbers of distracting documents (0, 3, 5, and 7).

read the caption

Table 6: Sample sizes in the constructed AntiLeak-Bench in the experiments.

🔼 This table presents the average word counts of samples in the constructed AntiLeak-Bench across different time periods (2022-01-01 to 2023-01-01 and 2023-05-01 to 2024-08-01), tasks (single-hop and multi-hop), and the number of distracting documents (0, 3, 5, and 7). The data is organized by time period, task type, and the number of distracting documents, allowing for an analysis of question complexity and context length across various experimental settings.

read the caption

Table 7: Average word counts of samples in the constructed AntiLeak-Bench in the experiments.

🔼 This table lists the release date and knowledge cutoff date for each of the large language models (LLMs) used in the study. The knowledge cutoff date refers to the point in time after which any newly generated knowledge is not included in the training dataset. An estimated cutoff date is marked with an asterisk.

read the caption

Table 8: Release dates and knowledge cutoff dates of LLMs. * means estimated time.