phenixace/OpenMolIns-xlarge

---
license: apache-2.0
language:
- en
dataset_info:
config_name: OpenMolIns-xlarge
size: 1200000
---

# OpenMolIns Instruction Tuning Dataset (XLarge)

Instruction tuning dataset for **Open-domain Natural Language-Driven Molecule Generation**, aligned with [S²-Bench (TOMG)](https://phenixace.github.io/tomgbench/).

This is the **xlarge** variant with **1,200,000** instruction–molecule pairs.

## Task Types

The dataset covers 9 molecular generation and optimization subtasks (aligned with S²-Bench configurations):

- **MolCustom_AtomNum**: Molecular customized generation by atom number
- **MolCustom_BondNum**: Molecular customized generation by bond number
- **MolCustom_FunctionalGroup**: Molecular customized generation by functional group
- **MolEdit_AddComponent**: Molecular editing – adding components
- **MolEdit_SubComponent**: Molecular editing – substituting components
- **MolEdit_DelComponent**: Molecular editing – deleting components
- **MolOpt_LogP**: Molecular optimization for LogP
- **MolOpt_MR**: Molecular optimization for MR
- **MolOpt_QED**: Molecular optimization for QED

## Dataset Structure

| Column | Description |
|-----------|--------------------------------------------|
| SubTask | One of: AtomNum, BondNum, FunctionalGroup, AddComponent, SubComponent, DelComponent, LogP, MR, QED |
| Instruction | Natural language instruction |
| molecule | Target molecule (SMILES) |

## Usage

```python
from datasets import load_dataset

# Load the xlarge training set
dataset = load_dataset("phenixace/OpenMolIns-xlarge")
# dataset["train"]: SubTask, Instruction, molecule
print(dataset["train"].num_rows) # 1200000
```

## OpenMolIns Variants

| Variant | # Instructions |
|---------|----------------|
| light | 4,500 |
| small | 18,000 |
| medium | 45,000 |
| large | 90,000 |
| xlarge | 1,200,000 |

## Evaluation

Models trained on OpenMolIns can be evaluated on [S²-Bench (TOMG)](https://huggingface.co/datasets/phenixace/S2-TOMG-Bench). See the [benchmark leaderboard](https://phenixace.github.io/tomgbench/) for results. The **OpenMolIns-xlarge** variant is used to train the top-performing model (Llama3.1-8B with OpenMolIns-xlarge) on the S²-Bench leaderboard.

## Citation

If you use this dataset, please cite:

```bibtex
@article{li2024speak,
title={Speak-to-Structure: Evaluating LLMs in Open-domain Natural Language-Driven Molecule Generation},
author={Li, Jiatong and Li, Junxian and Liu, Yunqing and Zheng, Changmeng and Wei, Xiaoyong and Zhou, Dongzhan and Li, Qing},
journal={arXiv preprint arXiv:2412.14642v3},
year={2024}
}
```

## Links

- [S²-Bench / TOMG Benchmark](https://phenixace.github.io/tomgbench/)
- [S2-TOMG-Bench GitHub](https://github.com/phenixace/S2-TOMG-Bench)
- [S²-Bench Dataset on Hugging Face](https://huggingface.co/datasets/phenixace/S2-TOMG-Bench)

---
许可证: Apache-2.0
语言:
- 英语
数据集信息:
配置名称: OpenMolIns-xlarge
数据规模: 1200000
---

# OpenMolIns 指令微调数据集（超大版）

本数据集为**开放域自然语言驱动的分子生成**专用指令微调数据集，与[S²-Bench (TOMG)](https://phenixace.github.io/tomgbench/)对齐。本数据集为**xlarge**变体，包含**1,200,000**条指令-分子配对数据。

## 任务类型
本数据集涵盖9类与S²-Bench配置对齐的分子生成与优化子任务：
- **MolCustom_AtomNum**: 基于原子数量的分子定制生成
- **MolCustom_BondNum**: 基于化学键数量的分子定制生成
- **MolCustom_FunctionalGroup**: 基于官能团的分子定制生成
- **MolEdit_AddComponent**: 分子编辑——添加组分
- **MolEdit_SubComponent**: 分子编辑——替换组分
- **MolEdit_DelComponent**: 分子编辑——删除组分
- **MolOpt_LogP**: 面向LogP的分子优化
- **MolOpt_MR**: 面向MR（摩尔折射率，Molar Refractivity）的分子优化
- **MolOpt_QED**: 面向QED（药物相似性定量估计，Quantitative Estimation of Drug-likeness）的分子优化

## 数据集结构
| 列名 | 描述 |
|-----------|--------------------------------------------|
| SubTask | 可选值包括：AtomNum、BondNum、FunctionalGroup、AddComponent、SubComponent、DelComponent、LogP、MR、QED |
| Instruction | 自然语言指令 |
| molecule | 目标分子（采用SMILES，即简化分子线性输入规范，"Simplified Molecular Input Line Entry System"格式） |

## 使用方法
python
from datasets import load_dataset

# 加载超大版训练集
dataset = load_dataset("phenixace/OpenMolIns-xlarge")
# dataset["train"] 包含 SubTask、Instruction、molecule 三列
print(dataset["train"].num_rows) # 1200000

## OpenMolIns 变体版本
| 变体版本 | 指令条数 |
|---------|----------------|
| light | 4,500 |
| small | 18,000 |
| medium | 45,000 |
| large | 90,000 |
| xlarge | 1,200,000 |

## 模型评估
基于OpenMolIns训练的模型可在[S²-Bench (TOMG)](https://huggingface.co/datasets/phenixace/S2-TOMG-Bench)上进行评估。可参考[基准测试排行榜](https://phenixace.github.io/tomgbench/)查看实验结果。S²-Bench排行榜上性能最优的模型（基于OpenMolIns超大版训练的Llama3.1-8B）即采用OpenMolIns-xlarge变体进行训练。

## 引用信息
若使用本数据集，请引用以下文献：
bibtex
@article{li2024speak,
title={Speak-to-Structure: Evaluating LLMs in Open-domain Natural Language-Driven Molecule Generation},
author={Li, Jiatong and Li, Junxian and Liu, Yunqing and Zheng, Changmeng and Wei, Xiaoyong and Zhou, Dongzhan and Li, Qing},
journal={arXiv preprint arXiv:2412.14642v3},
year={2024}
}

## 相关链接
- [S²-Bench / TOMG 基准测试](https://phenixace.github.io/tomgbench/)
- [S2-TOMG-Bench 开源代码库](https://github.com/phenixace/S2-TOMG-Bench)
- [Hugging Face 平台上的S²-Bench数据集](https://huggingface.co/datasets/phenixace/S2-TOMG-Bench)