Crisying/rgroup-generation-benchmark

Moved: This dataset is now part of the unified R-GroundBench release: https://huggingface.co/datasets/latent-recast-research/R-GroundBench

R-GROUNDBENCH: Generation QA Dataset

Dataset Description

R-GROUNDBENCH Generation QA is the open-ended generation track of the R-GROUNDBENCH benchmark. Given a Markush scaffold and a natural-language R-group substitution instruction, a model must directly generate the SMILES string of the edited molecule.

This track evaluates whether models can perform molecular editing without candidate options.

---

Benchmark Structure

The dataset has two difficulty splits:

Split	Instruction Style	Example
easy	Direct naming	Replace R1 with a 4-chlorophenyl group.
hard	Descriptive paraphrase	Replace R1 with a halogenated six-membered aromatic ring.

The easy-to-hard gap measures sensitivity to instruction style while keeping the output format fixed.

---

Dataset Statistics

Split	# Questions
easy	500
hard	500
Total	1,000

---

Data Fields

Each record contains:

difficulty             : str   easy or hard
original_esmiles       : str   Markush scaffold in E-SMILES format
original_image_path    : str   repo-relative path to the rendered scaffold image
instruction            : str   natural-language R-group substitution instruction
correct_smiles         : str   ground-truth edited molecule SMILES
scaffold_smiles_clean  : str   scaffold SMILES with annotation tokens removed
output_image_path      : str   repo-relative path to the ground-truth molecule image, when available

---

Evaluation Metrics

The public evaluator reports:

Metric	Description
Validity Rate	Fraction of outputs parseable by RDKit
Exact Match	Fraction where canonical SMILES equals the ground truth
Avg Tanimoto	Mean ECFP4 Tanimoto similarity to the ground truth
Scaffold Match	Fraction where the predicted molecule preserves the Markush scaffold

Models are expected to return:

<smiles>CCO</smiles>

---

Input Modalities

Mode	Input	Intended models
smi	E-SMILES text plus instruction	LLMs and VLMs
img	scaffold image plus instruction	VLMs only

---

E-SMILES Format

E-SMILES is a Markush-aware extension of standard SMILES. R-group positions are encoded with annotation tokens appended after a <sep> delimiter:

N1=CC(*)=C(*)N=C1*<sep><a>3:R[1]</a><a>5:R[2]</a><a>8:X</a>

E-SMILES strings should not be passed directly to RDKit without removing annotation tokens.

---

Usage

from datasets import load_dataset

dataset = load_dataset("Crisying/rgroup-generation-benchmark")
easy = dataset["easy"]

for record in easy:
    scaffold = record["original_esmiles"]
    instruction = record["instruction"]
    ground_truth = record["correct_smiles"]

---

Relationship to VQA Track

	VQA Track	Generation Track
Format	4-choice multiple choice	Open-ended SMILES generation
Splits	Easy / Medium / Hard	Easy / Hard
Primary metric	Accuracy	Exact Match, Tanimoto, Scaffold Match
Main capability	Candidate discrimination	Direct molecular editing

---

Citation

@article{rgroundbench2026,
  title  = {R-GROUNDBENCH: A Diagnostic Benchmark for R-Group Grounding in Markush Molecular Editing},
  author = {Anonymous},
  year   = {2026}
}

---

License

Released for research purposes. Source molecular records remain subject to their original licensing terms.