R-GroundBench

R-GroundBench is a diagnostic benchmark for evaluating R-group grounding in Markush molecular editing. It tests whether a model can connect R-group placeholders in a Markush scaffold to concrete substituents under natural-language instructions or property questions.

The release contains two tracks: multiple-choice visual question answering (VQA) and open-ended molecular generation. Every configuration contains a single test split. The benchmark records, their ordering, and the PNG assets are consolidated from the original VQA and Generation releases without regenerating questions or images.

Configurations and verified sizes

Track	Configuration	Rows
VQA	vqa_easy_basic	940
VQA	vqa_easy_advanced	634
VQA	vqa_medium_basic	940
VQA	vqa_medium_advanced	634
VQA	vqa_hard_basic	940
VQA	vqa_hard_advanced	634
Generation	generation_easy	500
Generation	generation_hard	500

The six VQA configurations contain 4,722 questions in total. The two Generation configurations contain 1,000 examples in total.

VQA track

The VQA track presents a Markush scaffold, an instruction or property question, and four candidate answers. The expected answer is A, B, C, or D.

• Basic tasks test direct R-group substitution: localize the requested R-group, interpret the substituent instruction, and select the edited molecule.
• Advanced tasks test property-guided candidate selection using physicochemical, biological-activity, or indication/mechanism information.
• Easy, Medium, and Hard vary distractor construction and instruction style. Easy Basic uses cross-scaffold distractors, Medium Basic uses same-scaffold distractors, and Hard Basic uses high-similarity same-scaffold distractors with descriptive instructions. Advanced configurations follow the corresponding direct-to-descriptive difficulty progression.
• NOTA robustness: Easy and Medium Basic include “None of the above” cases. The is_none_of_above, nova_answer_is_d, and nova_type fields preserve the associated annotations.

Where supported by a model, the benchmark can be evaluated in four modality settings:

Setting	Scaffold input	Candidate representation
img_img	Markush PNG	molecule PNGs
img_smi	Markush PNG	SMILES strings
smi_img	E-SMILES text	molecule PNGs
smi_smi	E-SMILES text	SMILES strings

Generation track

The Generation track supplies a Markush scaffold and a natural-language R-group substitution instruction. A model generates the complete edited molecule as SMILES without candidate options.

• generation_easy uses direct substituent naming.
• generation_hard uses descriptive paraphrases while retaining the same open-ended SMILES output format.

Loading

from datasets import load_dataset

dataset = load_dataset(
    "latent-recast-research/R-GroundBench",
    "vqa_easy_basic",
)
test = dataset["test"]

from datasets import load_dataset

dataset = load_dataset(
    "latent-recast-research/R-GroundBench",
    "generation_hard",
)
test = dataset["test"]

Important fields

VQA

• split: original difficulty/task split name.
• original_esmiles: Markush scaffold in E-SMILES format.
• original_image_path: repository-relative Markush scaffold PNG path.
• instruction, question, and property: task text and property metadata where applicable.
• correct_answer: correct option label (A, B, C, or D).
• correct_smiles: SMILES for the correct molecule.
• correct_prop_value and prop_values: property values where applicable.
• option_A_smiles through option_D_smiles: candidate representations.
• option_A_image through option_D_image: repository-relative candidate PNG paths. An empty option-D image is retained for applicable NOTA records.
• is_none_of_above, nova_answer_is_d, and nova_type: NOTA annotations.
• Hard configurations may include analysis metadata such as trick, n_rgroups, and is_same_group.

Generation

• difficulty: easy or hard.
• original_esmiles: input Markush scaffold in E-SMILES format.
• original_image_path: repository-relative scaffold PNG path.
• instruction: R-group substitution instruction.
• correct_smiles: ground-truth edited molecule SMILES.
• scaffold_smiles_clean: scaffold used by scaffold-preservation evaluation.
• output_image_path: repository-relative ground-truth molecule PNG path.
• Hard examples additionally retain trick and n_rgroups metadata.

E-SMILES is a Markush-aware extension of standard SMILES. Annotation tokens occur after a <sep> delimiter and should be removed before passing the scaffold to software that expects ordinary SMILES.

Image paths

Image columns are deliberately stored as strings, not Hugging Face Image features. Each non-empty value is a POSIX-style path relative to this dataset repository, under images/. The release retains source-grouped canonical paths where possible, stores content-identical PNGs only once, and preserves the original PNG bytes exactly.

To download a referenced asset:

from huggingface_hub import hf_hub_download

row = test[0]
local_png = hf_hub_download(
    repo_id="latent-recast-research/R-GroundBench",
    repo_type="dataset",
    filename=row["original_image_path"],
)

Evaluation overview

VQA is evaluated with exact multiple-choice accuracy after parsing the model response as A, B, C, or D.

Generation evaluation reports:

• validity: whether the generated output is parseable as a molecule;
• Exact Match: canonical-SMILES equality with the ground truth;
• Tanimoto similarity: ECFP4 fingerprint similarity to the ground truth;
• Scaffold Match: whether the generated molecule preserves the Markush scaffold under the benchmark’s scaffold-matching criterion.

Project links

• Paper: link to be added after the public paper URL is confirmed.
• Homepage: https://latent-recast-research.github.io/R-GroundBench/.
• GitHub: https://github.com/CrisYing/R-GROUNDBENCH
• modelscope:https://www.modelscope.cn/datasets/Latent-Recast-Research/R-GroundBench

License and source records

This benchmark is released for research use. Source molecular records remain subject to their original licensing terms.

Citation

Citation metadata will be added when the paper’s final public bibliographic information is confirmed.