MemoTrap

The MemoTrap dataset is a diagnostic dataset designed to assess whether language models (LMs) fall into memorization traps. These traps occur when LMs memorize specific examples from their training data instead of generalizing effectively. The dataset aims to probe the susceptibility of LMs to this phenomenon.

Here are some key details about the MemoTrap dataset:

  1. Purpose: MemoTrap serves as a benchmark for evaluating LMs' behavior in terms of memorization. It helps researchers and practitioners understand how well LMs generalize and whether they rely excessively on memorized examples.

  2. Contents: The dataset consists of four subtasks, each represented by a CSV file in the data folder:

    • Proverb Ending: Contains 860 instances related to English proverbs.
    • Proverb Translation: Includes 843 instances involving translations of proverbs.
    • Hate Speech Ending: Comprises 100 instances related to hate speech.
    • History of Science QA: Contains 736 instances related to questions about the history of science.

  3. Data Format:

    • Each data instance has three keys:
      • prompt: The input provided to the LM.
      • classes: A list of two candidate continuations for the prompt.
      • answer_index: The index of the desired continuation in the classes list.
    • Note that the classes field is a string version of a list and may contain escape characters.

  4. Sources: The data in MemoTrap is adapted from various sources:

    • English proverbs from Wikiquote⁴.
    • Historical science content from "The History of Science, Volume 1-5" by Ray Spangenburg and Diane Kit Moser⁴.

MemoTrap has been recognized as a prize-winning submission in the Inverse Scaling Challenge. Researchers use it to investigate how different LM sizes impact their susceptibility to memorization traps¹²³. 📚🔍

Source: Conversation with Bing, 3/17/2024 (1) undefined. https://en.wikiquote.org/wiki/English_proverbs. (2) GitHub - liujch1998/memo-trap. https://github.com/liujch1998/memo-trap. (3) arXiv:2305.14739v1 [cs.CL] 24 May 2023. https://arxiv.org/pdf/2305.14739.pdf. (4) The MemoTrap Dataset - GitHub. https://github.com/liujch1998/memo-trap/blob/d87b26abf4402b76d167d1326798d21d7ea2127e/README.md.