CB7-LLM-Inference
收藏资源简介:
CB7-LLM-Inference是一个用于评估大型语言模型(LLM)在化学宿主-客体结合亲和力预测任务中性能的数据集。该数据集聚焦于葫芦脲CB[7](cucurbituril)体系,涉及模型对结合常数对数(logKa)的预测。数据集包含三个子集:test子集包含63个未参与训练的“黄金”客体分子(held-out),用于最终测试;val子集包含125个校准客体分子,采用5折交叉验证(fold 0-4),用于模型校准与验证;SAMPL4子集包含14个来自SAMPL4宿主-客体挑战赛的客体分子,仅提供模型预测值,无真实测量值。每个样本包含以下字段:真实测量的logKa(true_logka,在SAMPL4子集中为null)、GPT-5.5模型的预测logKa及其原始响应文本、Claude Opus 4.8模型的预测logKa及其原始响应文本,以及交叉验证折叠标识(fold,val子集为0-4,其他为-1)。所有预测基于相同的V5宿主-客体提示模板和22个物理特征(如对接构象、空腔特征等),其中test/val子集特征来源于SupraBench/physics_feature,SAMPL4子集特征随挑战赛数据提供。该数据集适用于化学信息学、计算化学和LLM评估任务,特别是用于分析LLM在物理化学性质预测中的准确性与可靠性。
CB7-LLM-Inference is a dataset for evaluating the performance of large language models (LLMs) on chemical host-guest binding affinity prediction tasks. It focuses on the cucurbituril CB[7] system and involves predicting the logarithm of binding constants (logKa). The dataset consists of three subsets: the test subset contains 63 held-out gold guest molecules for final testing; the val subset contains 125 calibration guest molecules using 5-fold cross-validation (fold 0-4) for model calibration and validation; and the SAMPL4 subset includes 14 guest molecules from the SAMPL4 host-guest challenge, providing only model predictions without true measured values. Each sample includes the following fields: true measured logKa (true_logka, null in the SAMPL4 subset), predicted logKa and raw response text from the GPT-5.5 model, predicted logKa and raw response text from the Claude Opus 4.8 model, and a cross-validation fold identifier (fold, 0-4 for val subset, -1 for others). All predictions are based on the same V5 host-guest prompt template and 22 physical features (such as docking conformations, cavity features, etc.), with features for test/val subsets sourced from SupraBench/physics_feature and SAMPL4 subset features provided with the challenge data. The dataset is suitable for cheminformatics, computational chemistry, and LLM evaluation tasks, particularly for analyzing the accuracy and reliability of LLMs in predicting physicochemical properties.
CB7-LLM-Inference 数据集概述
基本信息
- 许可证:CC-BY-4.0
- 数据集名称:CB7-LLM-Inference
- 标签:化学、主客体化学、葫芦脲、结合亲和力、大语言模型
数据集内容
该数据集包含利用V5提示的大语言模型(LLM)预测的CB[7](葫芦[7]脲)logKa值与实测true_logka值的对比数据,涉及两种不同的大语言模型。
数据划分
| 划分 | 样本数 | 说明 |
|---|---|---|
| test | 63 | 保留的客分子(held-out),fold = -1 |
| val | 125 | 用于校准的客分子,包含5折交叉验证(fold 0-4) |
| SAMPL4 | 14 | SAMPL4主客体挑战中的客分子,无实测true_logka(值为null),fold = -1 |
数据列说明
| 列名 | 含义 |
|---|---|
true_logka |
实测logKa值(如无实测值则为null,例如SAMPL4数据) |
GPT-5.5 |
GPT-5.5模型预测的logKa值 |
GPT-5.5 Response |
GPT-5.5模型的原始响应 |
Opus 4.8 |
Claude Opus 4.8模型(中等思考模式)预测的logKa值 |
Opus 4.8 Response |
Opus 4.8模型的原始响应 |
fold |
val数据的交叉验证折数(0-4);test和SAMPL4数据值为-1 |
模型与特征说明
- 两种模型均使用相同的V5主客体提示(V5 host-guest prompt)以及相同的22个对接/空腔/机制特征(基于top-Boltzmann构象)
- test和val数据的特征来源于SupraBench/physics_feature
- SAMPL4数据的特征随SAMPL4客分子集合提供
- test/val数据的GPT-5.5通过Codex运行
- SAMPL4数据的两种模型均通过OpenRouter运行(openai/gpt-5.5, anthropic/claude-opus-4.8)




