MichiganNLP/language-energy-divide
收藏资源简介:
该数据集是论文《语言-能源鸿沟:测量多语言LLM推理的能源成本》的配套数据,发布了研究中使用的每种语言的能源测量数据和提示,以便研究人员无需重新运行完整的测量活动即可基于这些数据开展研究。数据集包含两个主要部分:一是能源测量结果(位于results/目录),包括Qwen3-8B在122种语言上的主要结果(如每令牌稳态能源、输出令牌总数、总GPU能源、准确率等),以及其他模型的跨模型研究、跨任务比较和批量大小扫描数据;二是提示数据(位于prompts/目录),包括122种Belebele语言的零样本思维链指令和引导词(经过机器翻译、后翻译质量控制和手动整理),以及机器翻译的GSM8K数学问题和LM-Arena开放聊天提示(带有质量分数)。数据集重点关注多语言LLM推理的能源效率差异,例如发现不同语言之间的每令牌能源差异高达8.3倍,总能源差异高达179倍,且能源消耗最高的语言准确率最低。数据基于ML.ENERGY基准(使用vLLM服务和Zeus库)在NVIDIA L40S和RTX 6000 Pro Blackwell硬件上测量,并提供了一个8语言子集(包括英语、中文、俄语、法语等高资源语言和南普什图语、提格里尼亚语、掸语、藏语等低资源语言)用于跨模型、跨GPU和跨任务实验。
This dataset is the supplementary data for the paper *Language-Energy Gap: Measuring the Energy Cost of Multilingual LLM Inference*. It publishes the energy measurement data and prompts for each language employed in the study, enabling researchers to leverage these data for their own research without rerunning the full measurement campaign. The dataset consists of two core components: 1. Energy measurement results (stored in the `results/` directory): including the main findings of Qwen3-8B across 122 languages, such as steady-state energy consumption per token, total number of output tokens, total GPU energy consumption, accuracy, and other metrics, as well as cross-model research data, cross-task comparison data, and batch size sweep data for additional models. 2. Prompt datasets (stored in the `prompts/` directory): including zero-shot chain-of-thought instructions and guiding prompts for all 122 Belebele languages, which underwent machine translation, post-translation quality control, and manual curation, as well as machine-translated GSM8K mathematics problems and LM-Arena open chat prompts with quality scores. This dataset centers on the disparities in energy efficiency across multilingual LLM inference scenarios. For example, it is observed that per-token energy consumption varies by up to 8.3 times across different languages, while total energy consumption varies by up to 179 times, with languages exhibiting the highest energy consumption also showing the lowest accuracy rates. The data was collected using the ML.ENERGY benchmark (with vLLM serving framework and the Zeus library) on NVIDIA L40S and RTX 6000 Pro Blackwell hardware. An 8-language subset is additionally provided, covering high-resource languages such as English, Chinese, Russian, and French, as well as low-resource languages including Southern Pashto, Tigrinya, Shan, and Tibetan, to support cross-model, cross-GPU, and cross-task experimental studies.




