rasynai/thermophysical-property-corpus

Rasyn Thermophysical Property Corpus

2,760,169 property values reported as results in the NIST ThermoML archive and in open-access papers, one row per value. Each row carries the value, its unit as written, the components of the material with their reported amounts where the source states them, the temperature and pressure where the source gives them, the reported uncertainty, the source DOI, title and licence, and the archive field or table cell the value was read from.

2,699,439 rows (97.8%) are a deterministic re-publication of the NIST ThermoML archive, one row per deposited value with its conditions, uncertainty and provenance. 60,730 rows (2.2%) were newly extracted from open-access papers by a language model; they are the lower-confidence part of this release and their measured accuracy is given below.

CC BY 4.0 covers Rasyn's compilation, meaning the selection, arrangement and added columns. It is not asserted over the underlying data, which are NIST public-domain records and CC BY papers under their own terms.

Built at Rasyn as training data for formulation-property models. Version unified-0.4.3, built 2026-09-06, first public release.

Terms used below. A record is the unit the pipeline extracted: for ThermoML one composition context of one data block of an archive file; for a paper one table together with one row or column header (one table can yield several records). A row is one measurement; a record has several. A document is one archive file or one paper.

from datasets import load_dataset
ds = load_dataset("rasynai/thermophysical-property-corpus")
ds["validation"][0]

import duckdb
duckdb.sql("SELECT property_name, count(*) FROM 'data/train-*.parquet' GROUP BY 1 ORDER BY 2 DESC")

Sources

source	What it is	Rows	Documents	Records
nist_thermoml_bulk	NIST Thermodynamics Research Center, ThermoML/Data Archive, NIST Public Data Repository doi:10.18434/mds2-2422, archive file ThermoML.v2020-09-30.tgz. Experimental data deposited with journal articles (J. Chem. Eng. Data, J. Chem. Thermodynamics, Fluid Phase Equilibria, Thermochimica Acta, Int. J. Thermophysics and others), publication years 2003-2019. Parsed from the archive's own JSON rendering of each XML file.	2,699,439	11,550	857,115
oa_ccby_documents	Table cells from 6,169 open-access documents (OpenAlex work types: article 5,970, preprint 148, data-paper 33, conference-paper 13, book-chapter 2, conference-abstract 2, review 1) with CC BY or public-domain terms, publication years 1988-2026, extracted by a language model and checked as described below.	60,730	6,169	28,772

One paper, 10.1021/acs.jced.6b00718, is present in both sources; its rows carry the licence label of the source they were read from, and the two sources date it one year apart.

Splits: train 2,223,990 rows (80.6%), validation 265,510 (9.6%), test 270,669 (9.8%). 150 MB of zstd Parquet in files of about one million rows, cut at record boundaries (test-0000.parquet, train-0000.parquet, train-0001.parquet, train-0002.parquet, validation-0000.parquet), row groups of 100,000 with a smaller last group per file.

PubChem experimental properties and a USPTO patent pilot were part of the working corpus and are not in this release. PubChem depositors set their own terms and several forbid commercial reuse; the patent pilot's property vocabulary was never curated. Both are marked review_required in the project's source registry.

Properties with the most rows (94 distinct names in all):

Property	Rows	Share
Mass density	832,109	30.1%
Mole fraction	364,187	13.2%
Viscosity	278,451	10.1%
Speed of sound	190,848	6.9%
Vapor or sublimation pressure	177,056	6.4%
Molar heat capacity at constant pressure	164,465	6.0%
Mass fraction	88,419	3.2%
Refractive index (Na D-line)	82,070	3.0%
Electrical conductivity	73,646	2.7%
Boiling temperature at pressure P	54,182	2.0%
Surface tension liquid-gas	47,469	1.7%
Excess molar enthalpy (molar enthalpy of mixing)	45,970	1.7%
Activity coefficient	44,979	1.6%
Thermal conductivity	33,889	1.2%
Solid-liquid equilibrium temperature	32,125	1.2%

Literature rows use five names; ThermoML rows use the archive's names.

Property	Rows	With a parsed temperature
Mass density	29,853	2,436 (8.2%)
Viscosity	18,547	3,318 (17.9%)
Thermal conductivity	8,161	615 (7.5%)
Kinematic viscosity	2,618	790 (30.2%)
Surface tension	1,551	244 (15.7%)

What a row looks like

A literature row with the unit quoted in the cell and a parsed temperature:

{
  "record_id": "lit:02645fa6a0830e23",
  "measurement_id": "jats-000-r1",
  "split": "train",
  "source": "oa_ccby_documents",
  "domain": "other",
  "subdomain": "literature:viscosity",
  "system_name": "crude oil",
  "n_components": 1,
  "components": "[{\"name\": \"crude oil\"}]",
  "component_inchikeys": null,
  "composition_disclosure": "partial",
  "property_name": "Mass density",
  "quantity": "density",
  "value": 0.85,
  "unit": "g/cm3",
  "value_raw_text": "0.85 [g/cm3]",
  "uncertainty_value": null,
  "uncertainty_unit": null,
  "uncertainty_text": null,
  "temperature_K": 295.15,
  "pressure_kPa": null,
  "temperature_quoted": "22 °C",
  "other_conditions": null,
  "test_method": null,
  "evidence_status": "reported_actual",
  "quality_tier": "bronze",
  "range_flag": null,
  "unit_inferred_from_header": false,
  "property_relabeled": false,
  "doi": "10.1038/s41598-022-13035-1",
  "document_id": "lit-10.1038/s41598-022-13035-1",
  "title": "Experimental characterization of colloidal silica gel for water conformance control in oil reservoirs",
  "publication_year": 2022,
  "license_label": "CC BY (version not recorded; see the paper's landing page); attribution via DOI",
  "license_url": "https://creativecommons.org/licenses/by/",
  "source_locator": "jats.table-wrap[0] id=Tab1 :: 0.85",
  "source_excerpt": "0.85",
  "quote_sha256": "1e181f0934d441445f03ff51c972ef44275b830c10a80401e53b27bf5baf327a",
  "extraction_method": "openai/gpt-oss-120b via vllm_table_extractor 0.2.0; verbatim verification by lit_table_extractor 0.1.0",
  "split_group_id": "sg:1c2935a2b67435a2",
  "tags": null,
  "schema_version": "0.1.1"
}

Columns

Column	Type	Meaning
record_id, measurement_id	string	Together the row key. measurement_id is the archive position (r12p1v0: row, property, value) for ThermoML and the table index plus row (jats-002-r7, pdftable-001-r3) for papers; it is unique within a record. 18 ThermoML records of one archive file carry .json inside record_id; the ids are kept stable.
split	string	train, validation or test.
source	string	nist_thermoml_bulk or oa_ccby_documents.
domain, subdomain	string	other for every row of this release. For literature rows subdomain is literature:<topic>, the keyword topic the paper was retrieved under, not the property of the row; ThermoML rows say thermophysical_mixture_property whether or not the record is a mixture.
system_name	string	All components joined with + for ThermoML (aniline + propan-1-ol); for literature, the table's row or column header as the model quoted it, which may itself contain +. Literature names are not curated: they name the system the way the table did, and in the sample measured below 19% were the wrong header for the cell or named another quantity.
n_components	int32	Number of components in the record. Always 1 for literature rows, where the whole system is one phrase.
components	string	JSON list, one object per component: name; inchikey (ThermoML); amount, basis, unit where the source states a fixed composition. basis is one of mol_fraction, mass_fraction, volume_fraction, molality (mol/kg), amount_concentration (mol/dm3), ratio (with unit mol/kg: amount of solute per mass of solution), mass_per_volume. An amount of 0.0 or 1.0 is an infinite-dilution or pure-component endpoint. Where a ThermoML block varies the composition, each composition is its own record, so composition never appears in other_conditions. For literature rows the list holds one object with the system phrase.
component_inchikeys	string	;-separated InChIKeys of the components, ThermoML only; null on 1,716 ThermoML rows whose compound had no identifier in the archive and on every literature row. Two components of a record can share an InChIKey (enantiomers).
composition_disclosure	string	full for ThermoML (unknown on the 1,716 rows without identifiers), partial for literature (the phrase names the system but not the amounts).
property_name	string	Property as named by the source.
quantity	string	The dimension the property was checked against (density, dynamic_viscosity, thermal_conductivity, temperature, ...); null for ThermoML properties the unit table does not cover.
value, unit	float64, string	The number and the unit string as written; never null. Units are not normalised. 1 on 636,977 ThermoML rows means dimensionless.
value_raw_text	string	Literature: the cell as quoted, then the quoted unit in brackets or [unit from context] when the model quoted none, then pipeline notes ([unit inferred from block header: X], [property relabeled from 'Viscosity': ...]). ThermoML: the archive's property label and unit followed by presentation: Direct value, X; the number itself is nPropValue inside source_excerpt.
uncertainty_value, uncertainty_unit, uncertainty_text	float64, string, string	ThermoML: the combined expanded uncertainty deposited with the value (2,641,013 rows, 97.8%), copied as deposited and not checked; 8,181 of them exceed ten times the value and carry uncertainty_exceeds_10x_value in range_flag. Literature: the ± x quoted in the cell (12,552 rows), unit as the row's.
temperature_K	float64	Temperature in kelvin: the ThermoML temperature constraint or variable (2,551,025 rows, 94.5%); for literature, temperature_quoted parsed when it is a single temperature with a unit inside 100 to 3500 K (7,403 rows, 12.2%; 305 more strings carry a unit but were not parsed).
pressure_kPa	float64	Pressure in kPa, ThermoML only (2,353,170 rows, 87.2%); null where a block carries two pressure entries.
temperature_quoted	string	Temperature string as quoted from a literature table (15,035 rows). Strings without a digit were nulled.
other_conditions	string	name=value pairs joined by ;. ThermoML: every constraint and variable of the block, keys Temperature, Pressure, Frequency, Wavelength, Mass density, Amount density, Specific volume, phase, written in that order where present, e.g. Pressure=101.325 kPa;Temperature=298.2 K;phase=Liquid. Literature: condition_as_quoted=<header text> for each header cell the model quoted; ; inside a header was replaced by ,.
test_method	string	Method as deposited (ThermoML only); the archive's placeholders (OTHER, UNKNOWN, ?, Unspecified) were nulled.
evidence_status	string	reported_actual on every row: the source reports the value as a result. The schema's other classes do not occur in this release.
quality_tier	string	Computed by the exporter's scorer, not reviewed. silver (2,697,722 rows, all ThermoML): structured provenance, composition and conditions, no preparation procedure. bronze (62,447 rows): every literature row, where the composition is a phrase, plus 1,717 ThermoML rows whose components could not be resolved to an identifier. There is no gold tier.
range_flag	string	outside_plausible_range when the SI value falls outside the unit table's range (temperature 0-2000 K, pressure 0-1e9 Pa, density 0.1-30,000 kg/m³, dynamic viscosity 1e-6-1e6 Pa·s, kinematic viscosity 1e-8-1 m²/s, thermal conductivity 1e-3-3000 W/m/K, surface tension 1e-4-1 N/m, specific heat 10-1e5 J/kg/K, molar heat capacity 0-1e5 J/mol/K, volumetric heat capacity 1e3-1e8 J/K/m³, molar energy ±5e6 J/mol, amount concentration 0-1e5 mol/m³, speed 10-1e4 m/s, pH -1-15) or, for diffusion coefficients and thermal diffusivity, outside 1e-15-1e-2 m²/s; uncertainty_exceeds_10x_value as described above. 12,408 rows carry a flag. Flagged rows are archive values kept as deposited; several are unit-scale errors in the source. Properties in other dimensions are not range-checked.
unit_inferred_from_header	bool	True on 5,460 literature rows whose unit was copied from the table header (step 6).
property_relabeled	bool	True on 2,618 literature rows relabelled from "Viscosity" to "Kinematic viscosity" because the unit is kinematic.
doi, title, publication_year	string, string, int32	The source publication; never null. Titles are the source's strings with markup removed.
document_id	string	ThermoML: the archive file's DOI; literature: lit- plus the DOI.
license_label, license_url	string	The terms the source was obtained under; three values occur, see "Licence and attribution".
source_locator	string	ThermoML: JSONPath of the value inside the archive's JSON file ($.PureOrMixtureData[0].NumValues[3].PropertyValue[0]). Literature: table id and the quoted cell, jats.table-wrap[2] id=T3 :: 0.998 for XML, pdf.tables[1] :: 0.998 or document.tables[1] :: 0.998 for converted PDFs (table index in conversion order; the same text can occur in more than one cell of a table). Non-breaking and zero-width spaces are normalised to plain spaces here and in value_raw_text; source_excerpt keeps the original characters.
source_excerpt, quote_sha256	string	The text the value was read from and the SHA-256 of that stored text. For ThermoML the excerpt is the PropertyValue node as canonical JSON (json.dumps(..., sort_keys=True)), which also carries the uncertainty.
extraction_method	string	deterministic thermoml_v2020_09_30 mapper or openai/gpt-oss-120b via vllm_table_extractor 0.2.0; verbatim verification by lit_table_extractor 0.1.0.
split_group_id	string	The group the split was decided on, see "Splits".
tags	string	;-separated. ThermoML: property_group:<archive group>. Literature: provenance:model-asserted on 1,097 rows whose proposal file lacked the extractor's own metadata (the run used no other model).
schema_version	string	0.1.1, the record schema the rows were flattened from.

How the ThermoML rows were made

Each archive file is one journal article's deposited data. The parser (thermoml_v2020_09_30, deterministic) reads every property value with its uncertainty, the block's constraints and variables, and the compounds with the identifiers the depositor gave, and writes one record per composition context of a data block with one measurement per value. Only values whose archive "presentation" is a direct value are in this release; 73,436 values presented as differences or ratios against a reference state were left out because they would read as property values under the plain property name. The archive sometimes repeats a state point inside a block (identical value, uncertainty and conditions): 7,620 rows are such repeats, distinguishable only by measurement_id and source_locator. ThermoML values and uncertainties are not checked beyond range_flag.

How the literature rows were made

1. Candidate documents came from an OpenAlex snapshot (works metadata, 2026), filtered to works whose open-access licence is CC BY, CC0 or public domain and whose title or abstract matched one of about thirty thermophysical-property topic phrases (thermal conductivity, viscosity, density, surface tension, flash point, permittivity, heat capacity, vapour pressure, solubility, nanofluids, ionic liquids and others; the topic is kept in subdomain). Keyword matching is loose, so off-topic documents were retrieved as well. Documents were fetched from PubMed Central or Europe PMC, or from the open-access location OpenAlex records for the work (publisher site or repository). Requests went only to locations OpenAlex records as open access; the pipeline did not itself test whether a location was behind a paywall.

2. PDFs were converted with Docling and JATS XML was parsed directly; every table was cut out with its caption and header.

3. openai/gpt-oss-120b (open weights, served with vLLM) was shown one table at a time and asked for rows of system, property, value, unit and conditions, each as a literal quote from the table. Two passes were run and their rows unioned, because output-length limits made single passes drop rows. This is the only non-deterministic step in the pipeline.

4. Every quoted field was matched character-for-character against the table text: 2,831,961 rows were verified, of which 11,489 named a property outside the seven asked for and were dropped, leaving 2,820,472 rows. (The proposal-level counts behind this figure are in the working logs, not in a published report.)

5. A number is only a property value if its unit says so. Each row had to carry a unit the pipeline's unit table recognises, with the dimension the property requires, giving an SI value inside a wide physical range. This removed most of the verified rows: 1,267,903 had no usable unit (none in the row or the table header, an ambiguous header, or a header unit giving an implausible value), 74,291 had no numeric value, 1,263,609 had a unit that was unrecognised, of the wrong dimension or gave an implausible value across two filter passes (the model had labelled particle sizes, moduli, temperatures and percentages as properties), and 26,345 unit-less "relative permittivity" values fell outside 1 to 100,000. 188,324 rows remained. "Viscosity" quoted in mm²/s or cSt was relabelled "Kinematic viscosity" from its unit.

6. Where a table stated its unit only in the header, the unit was copied from the header if exactly one unit family for that property appeared in the table text (5,460 rows in the release carry that note).

7. Two audits of the result showed that the checks above pass the assigned unit, not the cell, and pass any table whose numbers have the right shape. So the release preparation re-checks every document, table and cell:

   • Documents are dropped as a whole when their title matches a keyword list for clinical, human-subject and animal studies, or for simulation, prediction and modelling studies. The list is not exhaustive: 467 rows from 58 documents whose titles contain "in vitro", "in vivo" or "clinical" remain, most of them benchtop formulation studies. Tables are dropped when their caption or headers carry that vocabulary, tables credited to a third party (reprinted, reproduced, with permission, adapted or taken from), and tables whose text mentions simulation or fitted models (set aside for a scientific review that has not happened).
   • Each cell must parse as one number (sign, thousands separators, decimal point or comma, exponent, an optional ± uncertainty, parenthesised note, significance letter, unit token). Ranges, inequalities, citation markers, count-with-percentage cells and cells with other words are out; a unit token in the cell or its note must have the property's dimension. The value was re-derived from the cell text, which corrected 1,148 rows whose earlier value had lost a sign, an exponent or the digits before a thousands separator; d,ddd is read as a thousands separator unless the table uses decimal commas elsewhere.
   • Rows whose system name is not a material (no letters, numbers or ranges, placeholders and header words, property names, instrument or statistics words, clinical group labels, citations, figure references) are out, as are header-copied units where the system name or the quoted headers name another quantity.
   • Relative permittivity and Flash point were dropped as properties: with no unit, or with only a temperature unit, to check, the label had been attached to any number of the right shape (clinical statistics, glass transitions, inlet temperatures).
   • Per-property physical ranges were tightened (density 1 to 22,600 kg/m³, thermal conductivity 0.005 to 2,500 W/m/K, dynamic viscosity 1e-4 to 1e4 Pa·s, kinematic viscosity 0.1 to 5,000 mm²/s, surface tension 1 to 500 mN/m); units whose case makes the magnitude ambiguous (MPa·s, MN/m, W/MK) are out.
   • One cell of a table filed under two properties: all its rows are out. Two readings of one cell from the two passes: one kept, the one whose unit was quoted from the cell.

   Records removed at release preparation (whole documents or tables)	Records
   Document removed: clinical, human-subject or animal study (title)	5,652
   Document removed: simulation, prediction or modelling study (title)	6,978
   Table removed: clinical or human-subject vocabulary in caption or headers	2,639
   Table removed: third-party credit line	376
   Table set aside: possibly non-experimental, unreviewed	1,464
   Document removed: licence re-check failed or no licence signal	1,577

   Rows removed at release preparation (cell level)	Rows
   Relative permittivity rows	30,232
   Flash point rows	11,002
   Cell is not a single number, carries a citation, a count-with-percentage, other words, or a unit of another dimension	8,029
   Header-copied unit while the system name or headers name another quantity	3,768
   Outside the property's physical range	2,305
   Unit case makes the magnitude ambiguous	336
   System name is not a material	19,873
   One cell filed under two properties	83
   Second reading of a cell already kept	692

8. Every document's licence was checked against the OpenAlex snapshot worklist, the live OpenAlex API, the fetch log and, for PubMed Central documents, the licence code PubMed Central holds. A document was kept only if at least one of these signals existed and every one said CC BY, CC0 or public domain; NC, ND or SA clauses, unclear licences and disagreeing signals removed 1,577 records. These signals are not independent of one another: the snapshot worklist, the live API and the fetch log all derive from OpenAlex, which takes the licence from Crossref for anything outside PubMed Central, so only the PubMed Central code is a separate check and only for PubMed Central documents. The licence statement on the publisher's own page was never read. A further 4 documents rest on the snapshot worklist alone, the live API having no record or no licence for them.

What this guarantees: every literature value is a character-for-character quote of one number from a real table, the unit attached to the row has the dimension of the property, and where the cell states its own unit that unit agrees. What it does not guarantee: that the paper's number is right, that system_name and temperature_quoted are the correct row and column headers for that cell (the model chose them; they were checked to exist in the table, not to be the right headers), that the table reports a measurement rather than a calculation the text did not flag, or that a table is the paper's own work, since third-party credit lines were detected in captions and headers only and a table credited in a footnote or the body text may have survived. The literature rows have been measured once, on this release and after every filter above. 200 rows were sampled, 40 per property, chosen deterministically by hash, and each was judged against the text of the table it was quoted from. Weighted to the released literature rows:

Verdict	Share of literature rows
The named property of the named system	66%
Another quantity, or the wrong header for the cell	19%
A value the paper cited, calculated or listed as typical, not measured	15%

By property the correct share ran from 45% for thermal conductivity, where 16 of the 40 sampled rows were values cited from other papers, to 83% for kinematic viscosity. Roughly one literature row in three is therefore unreliable, and the failure is concentrated in header attribution and in reference tables rather than in transcription: every sampled value re-parsed exactly from its quoted cell.

Two caveats on that measurement. It is a sample, so at 40 rows per property the uncertainty on each per-property figure is around 15 percentage points. And the judging was done by the authors using a language model, not by a chemist. No row in this release has been reviewed by a chemist against its source paper. Treat literature rows as machine-extracted candidates rather than curated measurements; source_locator and source_excerpt let you check any row against its table.

Splits

Records are grouped by source document before splitting and a whole group goes to one split. Two records are in the same group if they share a DOI, a source-document hash (SHA-256 of the fetched document), a normalised title (exact, or token Jaccard of at least 0.8), or when their documents share at least ten identical measurements (system, property, value, unit, temperature, pressure) covering at least half of the smaller document, which catches a preprint and its journal version and a data set deposited under two papers (290 documents in 138 such clusters). There are 17,566 groups in all.

The split is not grouped by substance. Pure water measured in one paper can be in train and pure water measured in another paper in test: 4,826 test rows (1.8%) and 4,855 validation rows have an identical (system, property, value, unit) row in train, 2,043 test rows one that also agrees on temperature and pressure; almost all are ThermoML pure compounds, composition endpoints and mixtures re-reported in a second paper. An earlier build grouped by composition as well; because common systems appear in thousands of papers that rule chained most of the corpus into one group. For a substance-level holdout use component_inchikeys (ThermoML) or system_name.

Any group whose latest publication year is 2026 is frozen to test (3,986 rows in 450 groups, all literature); a group whose date could not be parsed would be frozen too, but none exists. The remaining groups are assigned by hashing: SHA-256 of split_group_id, first eight bytes as a fraction of 2^64, below 0.10 test, below 0.20 validation, else train. By group the split is 78.0% / 9.5% / 12.6%; by row 80.6% / 9.6% / 9.8%. By source the literature rows split 74.8% / 9.2% / 16.0% and the ThermoML rows 80.7% / 9.6% / 9.7%; 40.9% of the literature rows in test are 2026 documents. Test is therefore not an i.i.d. sample. Because groups are documents, a few property names and unit spellings occur only in one split.

Use validation for model selection and test for one final number. The split files are audited after writing: groups are recomputed from the written records and the build fails if any group appears in two splits.

Known limitations

• Coverage is uneven. ThermoML supplies most rows and is heavy on binary mixtures of common solvents and on pure compounds (2,156,939 ThermoML rows, 79.9%, are mixtures); commercial formulations, additive packages and nanofluids are a small minority. Literature rows lean towards 2018-2026.
• Units are the source's strings. Viscosity alone carries dozens of spellings (Pa*s, mPa·s, cP, P for poise). Map units case-sensitively and check magnitudes before converting. 225 distinct unit strings occur.
• ThermoML property names carry the archive's phrasing, so the same physical quantity can appear under several names (viscosity with and without a phase qualifier; Relative permittivity at zero frequency versus at various frequencies). property_name is a label, not an ontology; quantity is the nearest thing to one.
• Mole fraction, Mass fraction, Molality and Amount concentration rows are ThermoML measurements of an equilibrium composition (solubility, phase composition); the phase and the fixed variables are in other_conditions.
• ThermoML archive values that are unit-scale errors in the source (a diffusion coefficient of 4.7 m²/s, a density with an uncertainty of 10¹⁵ kg/m³) are flagged in range_flag, not removed. Temperatures of 0 K are extrapolated calorimetric endpoints; pressures reach 5.3 GPa.
• Literature conditions are strings. 7,403 literature rows have a parsed temperature; the rest have none in this dataset even if the paper gave one in the text.
• The same table cell can appear more than once in literature rows when the two extraction passes attributed it to different headers (12,723 rows share their cell with another row); source_locator identifies the cell. Readings that agree on headers, value and temperature were collapsed to one.
• The literature model chose system_name and temperature_quoted from the table's headers. They were checked to exist in the table, not to be the right headers for the cell; see the measured rates above.
• Author names are not recorded; attribution is by doi and title. Follow the DOI for the author list and the licence version (2.0, 2.5, 3.0 or 4.0 for CC BY papers; license_url is the unversioned CC BY directory and redirects to 4.0).
• Errata and retractions of source papers are not tracked.
• The extraction code, prompt, unit table and quarantined rows are not published with this release. The description above is the record of the procedure.

Licence and attribution

The compilation, its selection, arrangement and added columns are released under CC BY 4.0. That licence does not restrict the public-domain content and does not replace the attribution each CC BY paper requires. The underlying data keep their own terms:

• NIST ThermoML archive: works of NIST employees are not subject to copyright in the United States. The data may be subject to copyright outside the United States; NIST grants a non-exclusive, irrevocable, royalty-free worldwide right to publish, prepare derivative works and distribute them (https://www.nist.gov/open/license). NIST provides the data as is, makes no warranty of any kind, and does not endorse this compilation or any product. Please acknowledge NIST as the source. Cite NIST Thermodynamics Research Center, ThermoML/Data Archive, NIST Public Data Repository, doi:10.18434/mds2-2422 (archive v2020-09-30), the ThermoML software note Riccardi et al., J. Comput. Chem. 2022, 43(12), 879, doi:10.1002/jcc.26842, and the paper each record cites (doi, title).
• Open-access documents: CC BY of the version stated by the publisher, or public domain (US Government works), per license_label. Cite the document by doi and title.

Licence label	Rows	Documents
NIST ThermoML/Data Archive v2020-09-30 (doi:10.18434/mds2-2422); works of NIST employees are not subject to US copyright and NIST grants a worldwide royalty-free right to distribute; attribution to NIST and the cited paper required	2,699,439	11,550
CC BY (version not recorded; see the paper's landing page); attribution via DOI	60,727	6,167
Public domain (US Government work, as recorded by OpenAlex/PMC); attribution via DOI	3	2

Keep doi, title, license_label, license_url and source_locator if you redistribute a derivative. Source: National Institute of Standards and Technology, and the authors of the cited documents.

Maintenance

Rasyn maintains this dataset. Errors in specific rows, licence questions and takedown requests: open a discussion on the dataset page, or contact Rasyn directly if you would rather not raise it in public. Give a doi or a source_locator and we will identify and remove the affected rows.

Versions: the dataset is unified-0.4.3 (built 2026-09-06); schema_version 0.1.1 is the record schema; extraction_method names the parser or model per row. Changelog: unified-0.4.3, first public release.

Citation

@misc{rasyn_thermophysical_corpus_2026,
  title   = {Rasyn Thermophysical Property Corpus},
  author  = {{Rasyn}},
  year    = {2026},
  version = {unified-0.4.3},
  url     = {https://huggingface.co/datasets/rasynai/thermophysical-property-corpus},
  note    = {Hugging Face dataset, version unified-0.4.3}
}