ALD process parameter database v2.4 extracted from the full text of the literature

A database of atomic layer deposition (ALD) process parameters extracted from the full text of the scientific literature by an open-weight large language model, together with the manually annotated gold standards used to validate the extraction and a provenance table linking each value to a character span in its own source article. From 12,945 indexed records, the full text of 1,706 articles was retrieved and parsed; 656 of these never mention atomic layer deposition in their own text and were excluded. Eight process parameters (deposited material, precursor, co-reactant, deposition temperature, growth per cycle, chamber pressure, pulse time and purge time) were extracted from the remaining 1,050 articles, of which 597 yielded at least one value, giving 3,884 article-field rows. Splitting rows that hold more than one atomic value gives 5,565 individual values. What changed in version 2.4. The second annotation now covers the whole test set. Version 2.1 released a second, independent annotation of 25 of the 83 held-out articles; the remaining 58 were annotated afterwards by the same person, who did not see the first annotator's values. annotator_agreement_83articles.xlsx holds the two annotators' values and the verdict of the benchmark scoring rule for every cell of all 83 articles, and supersedes annotator_agreement_25articles.xlsx, which is kept so that the figures of the earlier versions stay reproducible. Of the 486 cells both annotators filled, 371 agree (76.3%); over the eight scored fields alone, 314 of 404 agree (77.7%); a further 91 cells were filled by one annotator only. Re-scoring the released predictions against the second annotation in place of the first gives F1 69.5% over the 421 cells it fills, against 66.0% on the frozen gold standard. The counts are deposited in analysis/. No other file changed: the database, the provenance table, the gold standard and Source_Data.xlsx are byte-for-byte those of version 2.3. What changed in version 2.3. Two things. The Figure 5 sheets of Source_Data.xlsx were renamed because that figure lost its single-bar panel and its panels are now (a)-(d): Fig5bc_per_article became Fig5b_per_article, Fig5d_by_material became Fig5c_by_material, and Fig5e_error_categories became Fig5d_error_categories. No cell of data changed. A new folder, analysis/, holds the computed results that the accompanying manuscript cites and that were not in any earlier version: the classification of 30 of the 453 articles that returned no value, the material-field score and the two-flag verification subset, the overlap with the community-curated ALD Database, and the 50 released values that two readers judged against their source articles. What changed in version 2.2. Literal null values returned by the model for three articles were removed, so the released table holds 3,884 rows rather than 3,886 and the provenance table 5,565 values rather than 5,571. An agreement column was added: every released value was compared with the value a second open-weight model returned for the same article and field, and the column records agree (2,797 cells, 72.0%), disagree (1,043, 26.9%) or one_only (44, 1.1%). No evaluation figure changes, because none of the three articles is in the held-out test set. What changed in version 2.1. The gold standard was corrected. In version 1.0 a block of rows carried the DOI of the row below it, an off-by-one introduced when two annotation files were merged in a spreadsheet, so 23 of the 84 held-out articles were scored against another article's annotated values. The pairing was restored from the annotator's per-DOI extraction notes and from the publisher-supplied title of each retrieved body, and the corrected file is released as ald_gold_standard_v1.1.xlsx with a row-by-row record of the correction. Two DOIs that did not resolve to a retrieved body were corrected, and one review article that reports no deposition of its own and carries no annotated value was removed from the test split, which is now 83 articles and 477 fields. The extraction output itself is unchanged: the error was in the answer key, not in the released data. Contents. ALD_DB_v2.0.xlsx holds the 3,884 article-field rows, one row per article per field; the file name is unchanged from version 2.0 because the released scripts open it by that name. provenance_spans.csv gives the character offsets and matched sentence for each individual value: 93.9% were located again in their own article and 75.5% matched verbatim or as a number carrying its unit; values that could not be located are marked rather than removed. ald_llm_fulltext_v2.xlsx is the frozen extraction output over all 1,050 articles. ald_gold_standard_v1.1.xlsx contains 477 manually annotated gold fields over 83 held-out articles, and relation_gold_104records.xlsx contains 104 human-verified condition records. annotator_agreement_25articles.xlsx holds a second, independent annotation of 25 of those articles beside the first annotator's values. split_manifest.json fixes the evaluation split. Source_Data.xlsx gives the numerical source data behind the figures, one sheet per figure panel. analysis/ holds the computed results described above. A recovered span shows that a value occurs at those offsets in that article. It is a claim a reader can check with the article and this table, verified by an automatic search released with the accompanying software. It is not a record of what the model attended to. Full texts of the source articles are covered by publisher licences and cannot be redistributed. Every record carries the article DOI so that the original text can be retrieved from the publisher. Accompanying software: Full-text extraction pipeline for ALD process parameters (v2.4), deposited separately under the MIT licence.