遇见数据集

GAIR/LIMR

收藏
Hugging Face2025-02-17 更新2025-04-12 收录
官方服务:

资源简介:

LIMR数据集是一个包含1,389个数学问题的精选数据集,用于挑战强化学习中数据量越大效果越好的假设。该数据集与Learning Impact Measurement (LIM)方法配合使用,能够自动化评估训练样本的有效性,减少对大量数据的需求,同时达到或超过使用完整数据集的训练效果。

The LIMR dataset is a curated collection of 1,389 mathematical questions designed to challenge the assumption that more data necessarily leads to better performance in reinforcement learning. Used in conjunction with the Learning Impact Measurement (LIM) methodology, this dataset enables automated evaluation of training sample effectiveness, reducing the need for large datasets while achieving or exceeding the performance of training with the full dataset.

提供机构:
GAIR
二维码
社区交流群
二维码
科研交流群
商业服务