lgilpin/uncertaintyEval

Dataset Card for Dataset Name

This dataset consists of three novel reasoning datasets. Each dataset contains a response prompt (a question to ask a user or an LLM) and a ground truth answer.

Dataset Details

This dataset consists of three tasks: counting, multihop, and factoid. Each row of the dataset consists of two columns: the prompt and the answer.
The counting dataset evaluates the ability to count dates. The multihop or RAG task evaluates the ability to retrieve information about open hours for CVS stores in California. The factoid evaluates the ability to retrieve information when there might not be an answer.

Dataset Description

• Dataset 1: Arithmetic (Counting):

• This dataset tests the ability to count days in a particular country with multiple arrival and departure dates. The ground truth answer is a single integer number. The dataset is generated for a set of 100 arrival and departure dates.

• Dataset 2: Multihop (RAGs): This dataset tests the ability to perform information retrieval to determine if there are 24/7 CVS pharmacies in a specified city. The ground truth answer is a single integer number indicating the number of 24-hour pharmacies, which could be 0. The dataset is generated for the 387 cities in California with a CVS store.

• Dataset 3: Factoid: This dataset tests the ability to retrieve facts where there may not be an answer. This dataset asks about the national sport for a specified country, which may not exist. The ground truth answer is a string indicating the national sport or "Nothing" if the national sport does not exist. This dataset is generated for the 249 countries in the world.

• Curated by: Nikhil Wani and Leilani Gilpin

• Funded by: Funding for this work was provided by Underwriters Laboratories Inc. through the Center for Advancing Safety of Machine Intelligence.

  Uses

  This dataset is designed to evaluate AI models across three key capabilities: arithmetic reasoning, multistep information retrieval, and fact-based knowledge queries. The arithmetic dataset involves calculating the total days spent in the U.S. based on simulated travel itineraries, testing logical date-based reasoning. The multihop dataset assesses a model’s ability to retrieve and aggregate information, such as identifying 24/7 CVS pharmacies in California cities through multiple steps. The factoid dataset evaluates the ability to retrieve factual information, including handling cases where no definitive answer exists, like determining a country's national sport. These datasets collectively serve as a benchmark for testing and improving AI systems’ reasoning, retrieval, and knowledge-handling skills.

  Direct Use

  This dataset is suited for benchmarking and improving AI systems in arithmetic reasoning, multistep information retrieval, and fact-based knowledge tasks. It can be used to train or evaluate models for the ability to handle uncertainty in tasks. It supports the development of retrieval-augmented systems and knowledge-based applications, including trivia systems and tools that handle "null answer" scenarios.

  Out-of-Scope Use

  This dataset is not suitable for tasks outside its specific domains of arithmetic reasoning, information retrieval, and fact-based question answering. It unsuitable for training or evaluating models on open-domain reasoning, ethical decision-making, or multimodal applications involving images or audio.

  Dataset Structure

  The dataset consists of two columns, the prompt and the true answer.

  Dataset Creation

  • Dataset 1: Arithmetic (Counting): This dataset tests the ability to count days in a particular country with multiple arrival and departure dates. The ground truth answer is a single integer number representing the number of days stayed in the US. This dataset is generated by creating four random dates to simulate a travel itinerary:

    • Start Date: A randomly selected date within the current year, marking the start of the trip.
    • First Departure: Set by adding a random number (up to 10) of days to the start date.
    • Second Arrival: Fixed to 40 days after the start date, ensuring it occurs after the first departure.
    • Second Departure: Determined by adding a random number (up to 10) of days to the second arrival date.

    The task is to calculate how many days were spent in the US, which is determined by summing by summing the difference between the first departure and start date, the difference between the second departure and second arrival, plus 2 (to account for fencepost errors.

  • Dataset 2: Multihop (RAGs): This dataset tests the ability to perform information retrieval to determine if there are 24/7 CVS pharmacies in a specified city. The ground truth answer is a single integer number indicating the number of 24-hour pharmacies, from 0 to 2. To generate the dataset, a list of 387 California cities with CVS stores was obtained by scraping the CVS store locator page: https://www.cvs.com/store-locator/cvs-pharmacy-locations/California/. For each city, we determined the number of 24-hour CVS pharmacies by scraping the pharmacy locations page for that city. Using BeautifulSoup, we checked each location's opening hours, adding one to the city count if a pharmacy operates 24/7.

  • Dataset 3: Factoid: This dataset tests the ability to retrieve facts where there may not be an answer. This dataset asks about the national sport for a specified country, which may not exist. The ground truth answer is a string indicating the national sport or "Nothing" if the national sport does not exist. This dataset is generated for the 249 countries in the world. To generate this dataset, we obtained a list of the 249 countries of the world using pycountry.countries. Then, we used the list of national sports from Wikipedia: https://en.wikipedia.org/wiki/National_sport and used BeautifulSoup to aggregate the national sport for each country (if it exists). For countries without a designated national sport, the value is set to "Nothing."

  Curation Rationale

  The curation rationale for this dataset is centered on creating targeted benchmarks to evaluate LLMs on uncertain tasks.Each dataset is designed to simulate realistic scenarios: the arithmetic dataset emphasizes logical reasoning with date-based calculations, reflecting common tasks like travel planning. The multihop dataset requires retrieving and aggregating structured information across multiple steps. The factoid dataset tests knowledge retrieval and handling of incomplete information, such as cases where answers may not exist.

  Source Data

  The source data comes from Wikipedia, and the CVS website.

  Data Collection and Processing

  Beautiful Soup was used to scrape information from the source data.

  Who are the source data producers?

  The source data for this dataset comes from publicly available resources and synthetic generation. The arithmetic dataset uses randomized date logic without external data. The multihop dataset relies on CVS's official store locator and pharmacy location pages. The factoid dataset draws from the pycountry.countries Python library for a list of countries and Wikipedia's national sports page for factual information.

  Personal and Sensitive Information

  There is no personal nor sensitive information.

  Bias, Risks, and Limitations

  Since the data is synthetically generated or scraped from specific online sources, it may not fully capture the diversity of real-world scenarios. For example, the arithmetic dataset's travel itineraries are based on random date generation, which may not reflect real-world travel patterns or demographic factors that could affect travel behavior. The multihop dataset is centered on CVS stores in California, may introduce regional bia. The factoid dataset relies on Wikipedia for national sports data, which may be incomplete or biased, as not all countries have a nationally recognized sport.