FineTuneBench
收藏资源简介:
FineTuneBench数据集由斯坦福大学创建,旨在评估商业化大型语言模型(LLMs)微调APIs在新知识注入和现有知识更新方面的效能。该数据集包含625个训练问题和1075个测试问题,覆盖新闻、虚构人物、医疗指南和代码更新四大领域。通过对比分析OpenAI和Google的五个前沿LLMs,包括GPT-4o和Gemini 1.5 Pro等,FineTuneBench揭示了微调服务在知识注入方面的显著不足。该数据集的创建过程基于最新新闻文章、虚构人物信息、医疗指南更新和代码框架变更,通过精心设计的问答对来测试模型的泛化能力。FineTuneBench的应用领域广泛,尤其适用于评估和改进LLMs在特定领域的知识更新和应用能力。
The FineTuneBench dataset was created by Stanford University, aiming to evaluate the efficacy of commercial large language model (LLM) fine-tuning APIs for new knowledge injection and existing knowledge updating. This dataset consists of 625 training questions and 1075 test questions, spanning four core domains: news, fictional characters, medical guidelines, and code updates. Through a comparative analysis of five state-of-the-art LLMs developed by OpenAI and Google, including GPT-4o and Gemini 1.5 Pro, FineTuneBench uncovers notable deficiencies in fine-tuning services for knowledge injection. The dataset is built upon recent news articles, fictional character data, updated medical guidelines, and code framework revisions, with meticulously crafted question-answer pairs to assess the models' generalization capabilities. FineTuneBench boasts broad application prospects, and is particularly valuable for evaluating and enhancing the knowledge updating and application capabilities of LLMs in specific domains.




