Safety Data Sheet Extraction Data for EHS and Chemical Compliance AI

What a usable SDS training corpus contains

A usable SDS corpus pairs each document with section spans and normalized field values, not just OCR text. In the US, OSHA's Hazard Communication Standard requires the information in Table D.1 under fixed section numbers and headings for sections 1 to 11 and 16, while sections 12 to 15 may be included but are not mandatory [1]. That rule is why section segmentation is a solvable task, and also why it is not trivial: optional sections are often present, sometimes stubbed, and sometimes missing entirely.

The section headings in Appendix D are 1 Identification; 2 Hazard(s) identification; 3 Composition/information on ingredients; 4 First-aid measures; 5 Fire-fighting measures; 6 Accidental release measures; 7 Handling and storage; 8 Exposure controls/personal protection; 9 Physical and chemical properties; 10 Stability and reactivity; 11 Toxicological information; 12 Ecological information; 13 Disposal considerations; 14 Transport information; 15 Regulatory information; and 16 Other information, including date of preparation or last revision [1]. Use those headings verbatim as your section label set, and add a separate label for content that sits outside any section, such as cover pages, distributor stamps and appended exposure scenarios.

Where a subheading has no relevant data, the SDS must say that no applicable information is available [1]. Your labels should distinguish "field present with value", "field explicitly marked not available" and "field absent". Collapsing these three states is a common source of false positives in compliance checks.

Fields that matter most for regulatory checks

The highest-value fields are the ones downstream rules consume: identifiers, classification, composition and exposure limits. A model that reads section 2 well but mangles section 3 percentages will fail the checks EHS teams actually run, such as flagging carcinogens above a threshold or matching ingredients against a restricted-substance list.

• Section 1: product name, product code, recommended use and restrictions, supplier name and address, emergency phone number.
• Section 2: hazard class and category pairs, signal word, hazard statements (H-codes), precautionary statements (P-codes), pictogram references.
• Section 3: chemical name, common name or synonym, CAS number, concentration as an exact value or range, and trade-secret withholding notes.
• Section 8: OSHA PEL, ACGIH TLV and other limits, with the limit type (TWA, STEL, ceiling), unit and the substance each applies to.
• Section 9: physical state, flash point, boiling point, vapor pressure, pH and similar values with units.
• Section 16: revision date and version number, which drive currency checks.

Composition tables are the hardest part. Ranges such as "10 - 30" or "<1", multi-row ingredient tables split across pages, CAS numbers with OCR errors in the check digit, and withheld identities all need explicit annotation rules. Validate CAS numbers with the check-digit algorithm at labeling time so corrupted values never become gold labels. For table-specific labeling conventions, see table structure recognition data.

Format variation: MSDS, regional SDS and multilingual sheets

Real SDS libraries mix current GHS-format sheets with legacy MSDS documents, regional variants and translations, so the corpus must sample all of them. Older pre-GHS MSDS layouts often use different section counts, orders and headings, which is exactly the variation that breaks template-based parsers.

GHS Annex 4 gives the international guidance for SDS preparation, but how it applies depends on what importing countries require [2]. The EU has its own REACH guidance on compiling SDS, separate from the UN annex [3], so an EU sheet and a US sheet for the same product can differ in sub-section content and regulatory references. Label the jurisdiction and format generation as document-level metadata, so you can measure accuracy per slice instead of hiding a weak slice in the average.

Multilingual coverage matters for global EHS platforms that index supplier sheets in German, Spanish, French or Chinese. Section numbers stay stable across languages, which makes them useful anchors, but heading text, decimal separators and unit conventions change. See multilingual business document data for script and language sampling patterns that also apply here.

Specification artifact: SDS extraction data request

A precise request makes supplier responses comparable and easier to assess. The template below shows the level of detail that separates a usable specification from a vague one.

Illustrative example: invented to show structure; it does not describe an available dataset.

An illustrative composition record in the output might look like this:

{"doc_id": "sds-000412", "section": 3, "row": 2,
 "chemical_name": "Ethanol", "cas": "64-17-5",
 "conc_low": 10, "conc_high": 30, "conc_unit": "% w/w",
 "trade_secret": false, "page": 2,
 "bbox": [72, 410, 540, 428], "value_state": "present"}

Rights and provenance questions specific to SDS

SDS are widely distributed, but wide distribution is not a training license. Manufacturers author the sheets; distributors, EHS software vendors and customers hold copies. A company that collected thousands of supplier sheets in its SDS management system holds copies of documents someone else wrote, so ask how the holder's right to share them for model training was established, and whether its own annotations and corrections are separately licensable.

The holder's added value is often the most useful part: verified field values entered into an EHS system, reviewer corrections, and links between sheet revisions. These pair well with documents, as described in documents paired with system-of-record entries, and correction logs follow the pattern in IDP human correction data. Section 1 and section 16 can also contain named preparers and phone numbers, which should be removed or replaced before delivery.

Document provenance and labeling decisions in machine-readable form. Croissant-RAI extends dataset metadata to cover life cycle and labeling documentation [4], and a dataset card records intended uses and known biases [5], such as over-representation of one manufacturer's template.

Evaluating SDS extraction quality

Evaluate at the field level, sliced by format generation, jurisdiction, language and scan quality. Report section-boundary accuracy separately from field extraction, because a model can find section 8 correctly and still assign a STEL value to the wrong substance.

Hold out whole manufacturers, not random pages, so the test set measures generalization to unseen templates. Score composition rows as tuples (name, CAS, low, high, unit) rather than as independent fields, and track how often "not available" states are predicted as values. For gold-label construction and scoring conventions, see document extraction evaluation sets and key-value extraction labels.

How SourceX approaches SDS sourcing requests

SourceX sources operational datasets from US companies on request, and documents are among the kinds of data it looks for. It does not hold SDS in stock, and a request does not guarantee a match. You describe the documents and labels you need, not the businesses that might hold them, and every release is approved by the supplying company.

Each dataset is rights-reviewed for ownership and consents, and personal details such as names, emails and phone numbers are removed or replaced before delivery, with the method recorded and a sample checked. No method is perfect. If you are scoping SDS extraction work, describe your requirements on the SourceX buyers page. For broader document categories, see the Document AI data hub and SourceX's technical documents overview.

Request safety data sheet extraction data

SourceX looks for US businesses that hold the SDS documents and labels you describe, then handles the commercial process from assessment through a license that defines records, uses, term and delivery. Nothing is contracted until a supplier agrees. Start an SDS data request.