Dataset for "Large language models in materials science and the need for open-source approaches"

Supporting dataset for: “"Large language models in materials science and the need for open-source approaches”, Fengxu Yang and Jack D. Evans, 2025 This repository provides benchmarking and fine-tuning tools for applying LLMs to metal-organic framework (MOF) synthesis analysis. Key components: Extraction: Performance evaluation of 6 LLMs (GLM-4.5, Qwen series, DeepSeek) on extracting synthesis conditions (precursors, solvents, temperature, time, etc.) from text Prediction: LoRA fine-tuning pipeline for predicting synthesis parameters from unstructured text Datasets: Training and test sets in JSONL format with structured synthesis information Results: Comprehensive benchmarks including accuracy, inference speed, and VRAM requirements Reproduces and extends methodologies from MOF ChemUnity and L2M3 projects.本数据集为论文《大型语言模型在材料科学中的应用及开源路径的必要性》（Large language models in materials science and the need for open-source approaches）的配套数据集，作者为杨凤旭与杰克·D·埃文斯，2025年。 本仓库提供了将大语言模型（LLM）应用于金属有机骨架（MOF）合成分析的基准测试与微调工具。 关键组成部分： - 信息抽取：针对6款大语言模型（GLM-4.5、通义千问（Qwen）系列、DeepSeek）开展性能评估，测试其从文本中抽取合成条件（前驱体、溶剂、温度、反应时长等）的能力； - 预测建模：搭建LoRA（低秩适配）微调流水线，实现从非结构化文本中预测合成参数； - 数据集：采用JSONL格式存储的结构化合成信息训练集与测试集； - 测试结果：涵盖准确率、推理速度与显存（VRAM）占用需求的全方位基准测试结果。 本工作复现并拓展了MOF ChemUnity与L2M3项目的研究方法。