The rapidly expanding body of chemical literature contains physical data scattered across unstructured text and complex tables in thousands of publications. Translating this information into machine-readable formats is essential for training data-driven chemical models. The lack of uncertainty quantification in computationally extracted datasets makes manual extraction and curation still necessary. We here address this challenge and present an uncertainty-quantified data extraction pipeline. Our fully automated pipeline utilizes large language models (LLMs) to extract complex tabular and textual chemical data directly from scientific PDFs. We systematically assess several modern LLMs and document-to-text conversion frameworks. We benchmark the uncertainty in extracted data by comparison of our computationally extracted databases with large, existing databases on solubility that are manually curated. Our pipeline achieves between 93 and 79% of the literature contained data, with near-zero median log deviations for structurally matched rows and residual high-magnitude errors concentrated in a small subset of OCR, table-structure, unit-scaling, or reference-dataset discrepancies. We showcase the utility of our pipeline by deploying it to extract data from over 2,000 articles and generate BigMixSolDB, a comprehensive database on solubility in complex mixtures. BigMixSolDB comprises 280,273 solubility entries spanning single, binary, and ternary solvent systems. Our results demonstrate how integrated LLM-based pipelines can be used for literature data extraction that is comparable with manual curation. We envision that these frameworks can be used for large-scale, accurate extraction of literature data into datasets for machine learning applications
Andrei Voinea, Anna C. M. Thöni, Elija Veenman et al.· ChemRxiv· 0 citations
Supplementary material for the publication "BigMixSolDB: Extraction of a solubility database in solvent mixtures with an uncertainty-quantified large language model-based pipeline". This Zenodo record contains the data and supporting materials for BigMixSolDB, a workflow for extracting experimental solubility data from scientific PDFs. The archive provides the final standardized datasets generated with Docling- and vision-language-model (VLM)-based document-conversion pipelines, curated molecule dictionaries, corpus-screening and deduplication records, model-evaluation outputs, extraction prompts and requests, per-article extraction outputs, and a source-code snapshot. Together, these materials support inspection and reproduction of the extraction, post-processing, filtering, and evaluation procedures described in the accompanying publication. The Docling-derived dataset is the primary BigMixSolDB release because this pipeline achieved the strongest structural coverage and chemical-identity matching in the manuscript benchmarks. The VLM-derived dataset is provided as a complementary output for methodological comparison. Files BigMixSolDB_code.zip: Snapshot of the source code and workflow used to generate, standardize, filter, deduplicate, and evaluate the deposited datasets. The actively maintained repository is also available on GitHub. bigmixsoldb_docling.csv: Final standardized and deduplicated Docling-derived BigMixSolDB dataset. This is the primary curated dataset described in the publication and contains 280,161 retained solubility records from 2,150 represented source DOIs. bigmixsoldb_vlm.csv: Final standardized and deduplicated VLM-derived dataset, containing 231,936 retained records from 2,067 source DOIs. This file is supplied as a complementary pipeline output and methodological comparison rather than as the primary BigMixSolDB release. bigmixsoldb_docling_duplicates.csv and bigmixsoldb_vlm_duplicates.csv: Reports of duplicate groups identified during within- and cross-DOI deduplication. The procedure accounts for normalized pressure, equivalent pressures within 2%, and constrained differences in reported decimal precision. bigmixsoldb_docling_unfiltered.csv and bigmixsoldb_vlm_unfiltered.csv: Unfiltered datasets resulting from parsing the YAML outputs generated by the large language model. doi_counts.csv: Per-DOI comparison of retained coverage in BigMixSolDB against BigSolDB 2.0 and MixtureSolDB. dois.csv: List of publication DOIs retrieved during construction of the BigMixSolDB corpus. excluded_solute_names.txt and excluded_solvent_names.txt: Curated lists of compound names excluded by the molecular-scope filters. extracted_docling.zip and extracted_vlm.zip: Archives of the per-article structured extraction outputs generated with the Docling- and VLM-based pipelines, respectively. These broader extraction outputs are retained to support auditing and alternative downstream filtering. model_evaluation.zip: Validation- and test-set model-evaluation materials used to calculate the precision, recall, and F1 values reported in Table 1 of the publication. The archive contains model outputs, standardized reference and prediction CSV files, per-document error results, aggregate error summaries, and the evaluation prompts for the evaluated proprietary, open-weight, and fine-tuned models. name_to_smiles_docling.json and name_to_smiles_vlm.json: Curated molecule dictionaries used to assign and validate molecular structures in the respective pipelines. These files preserve compound-level provenance and permit users to apply alternative scope criteria. prompt_extraction.txt: Prompt template used to extract structured solubility records from converted document content. prompt_training.txt: Prompt template used in model fine-tuning. removed_dois.txt: Source DOIs excluded during article-level screening because they did not report in-scope experimental molecular solubility or solid-liquid-equilibrium data. requests_docling.jsonl and requests_vlm.jsonl: Batch request records for the Docling- and VLM-based extraction runs, respectively. The final CSV files report temperature in Kelvin and reported pressure in Pascal. Solubility is expressed in fraction-based units, and solvent composition is expressed as mole or mass fraction. A blank pressure denotes that pressure was not explicitly available to the extraction pipeline or present in the source article. The "Requires Review" field is a conservative indicator of potentially difficult source formatting or extraction ambiguity and should not be interpreted as evidence that a record is incorrect. Users should apply additional compound- or task-specific filters where required by their downstream application. Changelog Version 4 This version updates the two final datasets and their respective name-to-SMILES dictionaries. - Canonicalized the curated SMILES in both extraction tracks before generating the deposited CSV files. This changed 141 entries in the Docling name-to-SMILES mapping and 122 entries in the VLM mapping. The revised Docling dataset now has 334 unique non-empty values in both `Solvent 1` and `SMILES_Solvent_1`. - Added normalized compound display names for downstream name-based grouping while retaining the original extracted `Compound Name` values for provenance. - Audited the ternary subset and removed records in which phase-forming solids had been represented as inert solvents or the solute fraction had been folded incorrectly into an inferred third-solvent fraction. The Docling dataset removes 67 rows from DOI `10.1016/j.jct.2017.03.008`, 30 affected rows from DOI `10.1016/j.jct.2025.107514`, and 15 affected rows from DOI `10.1021/je201047m`, for a total of 112 removed rows. The corresponding VLM dataset removes 167 rows across the same three sources. Unaffected records from the latter two sources remain. The final Docling-derived release contains 280,161 records from 2,150 represented source DOIs and 1,668 unique solutes. It comprises 115,203 single-solvent, 161,921 binary-solvent, and 3,037 ternary-solvent records. The binary subset spans 834 unique solvent pairs and 3,352 unique solute-solvent mixture systems. The ternary subset spans 120 unique solvent triples and 146 unique solute-solvent mixture systems. Thus, the binary and ternary subsets together contain 3,498 unique solute-solvent mixture systems. The complementary final VLM-derived release contains 231,936 records from 2,067 represented source DOIs. Version 3 This version updates the deposited data and supporting materials to correspond to the final filtering, deduplication, nomenclature, and evaluation procedures reported in the manuscript. Regenerated both standardized datasets after source-level screening and revised molecular-scope filtering. The curated release is restricted to experimental solubility and solid-liquid-equilibrium measurements for molecular solutes in molecular solvent systems. Papers reporting only liquid-liquid equilibrium, partitioning, extraction efficiency, binodal or cloud-point data, purely computational results, or no original experimental solubility data were excluded. Removed 247 out-of-scope source DOIs before extraction and excluded 17,006 Docling-derived rows under the molecular-scope criteria, including inorganic salt systems and ionic compounds used as solvent components. Added pressure- and precision-aware within- and cross-DOI deduplication. This reduced the filtered Docling dataset from 278,688 to 274,885 rows by removing 4,204 duplicate rows, of which 4,105 were associated with cross-DOI groups and 99 with within-DOI groups. Standardized solvent text fields to one canonical display name per structure-defined solvent identity. Curated common names are used where available and deterministic frequency-based names otherwise. Updated the curated molecule JSON files. The revised records include source DOI provenance, solute or solvent role, molecule type, inclusion status, and structure-retrieval and human-validation metadata. Added the model-evaluation files used to generate the model-comparison results in Table 1, including the validation and test outputs, standardized reference/prediction tables, and aggregate error summaries. Uploaded the duplicate reports and updated DOI coverage information, exclusion lists, and source-screening records to match the regenerated datasets. In this version, the Docling dataset contains 280,273 records from 2,151 represented source DOIs. This comprises of 115,203 single-solvent, 161,921 binary-solvent, and 3,149 ternary-solvent records, with 1,676 unique solutes, 345 unique solvents, 839 binary solvent systems, and 124 ternary solvent systems. Version 2 This version updated the Zenodo repository to match the post-processed and filtered BigMixSolDB release described in the preceding manuscript revision. Replaced the preliminary dataset files with the final filtered Docling-derived BigMixSolDB release containing 299,211 retained solubility entries from 2,301 source DOIs. Corrected the `review_required` flag, which had previously been incorrectly set to `False` for all rows in the deposited dataset. Standardized retained temperature, pressure, solubility, and solvent-composition units. Temperatures are reported in Kelvin, reported pressures in Pascal, solubilities and solvent compositions in fraction-based units. Manually curated solute and solvent name-to-SMILES mappings, corrected erroneous molecular identifiers, resolved ambiguous assignments, and removed rows with missing or invalid required SMILES. The raw extracted dataset can be found in `bigmixsoldb_docling_unfiltered.csv`. A list of molecules that were omitted in our dataset can be found in `excluded_solute_names.txt` and `excluded_solvents.txt`. Post-processing steps and filtering the molecules found in the aforementioned TXT files gives the final dataset present in `bigmixsoldb_docling.csv`. Applied final s
Andrei Voinea, Chiara Thöni, Elija Veenman et al.· Zenodo (CERN European Organi...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
