A Large-Scale Dataset for Molecular Structure-Language Description via a Rule-Regularized Method

Researchers have developed an automated framework to generate a large-scale dataset of 163,000 molecule-description pairs by combining rule-based chemical nomenclature parsing with LLM guidance, achieving 98.6% precision in aligning molecular structures with natural language descriptions. This addresses a critical bottleneck in training language models for chemistry applications where manual annotation is prohibitively expensive.

This research tackles a fundamental challenge in applying large language models to chemistry: the scarcity of high-quality, structure-grounded datasets. By automating the annotation process through a rule-regularized framework that extends existing IUPAC nomenclature parsing, the team circumvents the traditional bottleneck of manual human annotation while maintaining exceptional precision. The approach demonstrates that combining deterministic rule-based systems with neural language models can produce reliable training data at scale.

The work emerges from growing recognition that LLMs require domain-specific, accurately aligned datasets to reason effectively about downstream tasks. Chemistry presents unique challenges because molecular structure—the primary determinant of function—must be preserved with mathematical precision in natural language descriptions. Previous attempts at structure-language alignment have either relied on expensive human curation or sacrificed accuracy for scale.

This dataset represents a significant infrastructure contribution to the emerging field of AI for chemistry. Access to 163,000 validated molecule-description pairs enables researchers to develop and fine-tune models capable of understanding chemical properties from text, supporting applications ranging from drug discovery to materials science. The 98.6% precision rate, validated through both LLM and expert human evaluation, provides confidence in dataset quality.

The open-source release through GitHub and Hugging Face democratizes access to this resource. Future development will likely focus on extending these methods to other complex molecular representations and exploring how models trained on this dataset perform on real-world chemical reasoning tasks. The framework's automation approach may also provide a template for curating large-scale datasets in other technical domains requiring precise structural information.

• →Automated annotation framework achieves 98.6% precision in generating 163,000 molecule-language pairs without manual curation.
• →Rule-regularized IUPAC nomenclature parser combined with LLMs produces structured XML metadata that preserves complete molecular details.
• →Open-source dataset release accelerates development of language models for chemistry and drug discovery applications.
• →Approach demonstrates that hybrid rule-based and neural methods can solve domain-specific dataset curation challenges at scale.
• →Dataset provides reliable foundation for aligning chemical structure understanding with natural language reasoning in LLMs.