Datasets:

phenixace
/

S2-TOMG-Bench

license: apache-2.0
language:
  - en
configs:
  - config_name: MolCustom_AtomNum
    data_files:
      - split: test
        path: MolCustom_AtomNum.csv
  - config_name: MolCustom_BondNum
    data_files:
      - split: test
        path: MolCustom_BondNum.csv
  - config_name: MolCustom_FunctionalGroup
    data_files:
      - split: test
        path: MolCustom_FunctionalGroup.csv
  - config_name: MolEdit_AddComponent
    data_files:
      - split: test
        path: MolEdit_AddComponent.csv
  - config_name: MolEdit_SubComponent
    data_files:
      - split: test
        path: MolEdit_SubComponent.csv
  - config_name: MolEdit_DelComponent
    data_files:
      - split: test
        path: MolEdit_DelComponent.csv
  - config_name: MolOpt_LogP
    data_files:
      - split: test
        path: MolOpt_LogP.csv
  - config_name: MolOpt_MR
    data_files:
      - split: test
        path: MolOpt_MR.csv
  - config_name: MolOpt_QED
    data_files:
      - split: test
        path: MolOpt_QED.csv

S^2-Bench Dataset (TOMG) (full version, 45k entries)

Official Huggingface Datasets for S^2-Bench: "Speak-to-Structure: Evaluating LLMs in Open-domain Natural Language-Driven Molecule Generation"

Please refer to our Github Repo for more usage and useful information.

Configurations

Each configuration represents a different task:

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

Usage

from datasets import load_dataset

# Load a specific configuration
dataset = load_dataset("phenixace/S2-TOMG-Bench", "MolCustom_AtomNum")

# Or load all configurations
configs = ["MolCustom_AtomNum", "MolCustom_BondNum", "MolCustom_FunctionalGroup",
           "MolEdit_AddComponent", "MolEdit_SubComponent", "MolEdit_DelComponent",
           "MolOpt_LogP", "MolOpt_MR", "MolOpt_QED"]

datasets = {config: load_dataset("phenixace/S2-TOMG-Bench", config) for config in configs}

Citation

If you use our data, please cite us in the format below:

@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}
}

Current Leaderboard

Rank	Model	#Parameters (B)	$\overline{S!R}$ (%)	$\overline{W!S!R} (%)$
1	Llama3.1-8B (OpenMolIns-xlarge)	8	58.79	39.33
2	Claude-3.5	-	51.10	35.92
3	Gemini-1.5-pro	-	52.25	34.80
4	GPT-4-turbo	-	50.74	34.23
5	GPT-4o	-	49.08	32.29
6	Claude-3	-	46.14	30.47
7	Llama3.1-8B (OpenMolIns-large)	8	43.1	27.22
8	Galactica-125M (OpenMolIns-xlarge)	0.125	44.48	25.73
9	Llama3-70B-Instruct (Int4)	70	38.54	23.93
10	Galactica-125M (OpenMolIns-large)	0.125	39.28	23.42
11	Galactica-125M (OpenMolIns-medium)	0.125	34.54	19.89
12	GPT-3.5-turbo	-	28.93	18.58
13	Galactica-125M (OpenMolIns-small)	0.125	24.17	15.18
14	Gemma3-12B	12	26.28	15.00
15	Deepseek-R1-distill-Qwen-7B	7	25.07	14.61
16	Llama3.1-8B-Instruct	8	26.26	14.09
17	Llama3-8B-Instruct	8	26.40	13.75
18	chatglm-9B	9	18.50	13.13(7)
19	Galactica-125M (OpenMolIns-light)	0.125	20.95	13.13(6)
20	ChemDFM-v1.5-8B	8	18.24	12.07
21	ChemLLM-20B	20	16.23	9.76
22	Llama3.2-1B (OpenMolIns-large)	1	14.11	8.10
23	yi-1.5-9B	9	14.10	7.32
24	Mistral-7B-Instruct-v0.2	7	11.17	4.81
25	BioT5-base	0.25	24.19	4.21
26	MolT5-large	0.78	23.11	2.89
27	Llama3.1-1B-Instruct	1	3.95	1.99
28	MolT5-base	0.25	11.11	1.30(0)
29	MolT5-small	0.08	11.55	1.29(9)
30	Qwen2-7B-Instruct	7	0.18	0.15

OpenMolIns

The instruction tuning datasets are also available at Hugging Face Datasets: