vallydia/ingredient-evidence-register

Vallydia Ingredient-Evidence Register

An open, evidence-graded reference dataset of 85 cosmetic and research ingredients — peptides, small molecules, proteins and blends — published by Vallydia under CC-BY-4.0.

Most ingredient databases tell you what a compound is. This one tells you how good the evidence actually is, outcome by outcome, and says so honestly when the answer is "we don't know" or "it was tested and it failed".

• Per-outcome A–F evidence grades — a compound can be B for one outcome and F for another. 456 graded outcome rows.
• DOI-verified citations — 117 citations resolved to a DOI with a verified title. This is the credibility spine of the dataset.
• Machine-readable cosmetic-claims maps — for each cosmetic ingredient, the claims that may be made and the claims that must not be. 246 claim rows.
• Legal status by region (INT / EU / US / UK) and WADA prohibited status.
• Chemical identifiers — CAS, PubChem CID, InChIKey, SMILES — each with an explicit confidence level and provenance. Where no confident match exists, the field is blank on purpose.
• Molecular structure images rendered from those identifiers, correct by construction.

Scope, and what is deliberately absent

This is an appearance-and-evidence reference. It contains no dosing, no administration or reconstitution information, and no commerce data — not in any field, and none is introduced by the generator. Research and injectable compounds (GLP-1 analogs, BPC-157, melanotan, and so on) are included as reference entries with their honest grades and legal status, never framed as usable or obtainable.

This is not medical advice. The grades are Vallydia's own evidence assessment, not third-party certification or peer review (reviewer_status is unassigned on every row, and we say so rather than implying a review that did not happen).

Dataset summary

Compounds	85
Classes	peptide (53), small_molecule (23), blend (5), protein (3), polysaccharide (1)
Graded outcome rows	456 (97 intentionally ungraded — see below)
Cosmetic ingredients	21, with 246 permitted/forbidden claim rows
Legal-status rows	132
Cited sources	373, of which 117 are DOI-verified across 50 compounds
WADA-prohibited	19
Confident chemical identifiers	64 of 85 (21 honestly blank)
Structure images	42 2D structures + 22 large-molecule data cards

Supported tasks

Evidence-grade classification, evidence-grounded question answering, regulatory-claim checking (which claims are permissible for a given cosmetic ingredient), and retrieval-augmented generation over data/corpus.jsonl.

Languages

English (en).

Dataset structure

Canonical, lossless

• data/compounds.jsonl — one JSON object per compound with the full nested structure, including the parsed Markdown body (body_sections) and a raw copy (body_markdown). This is the master file and the primary RAG-ingestion format.
• data/compounds.json — the same content as a single JSON array.

Flat table

• data/compounds.csv — one row per compound: scalar core plus |-joined lists, and best_outcome_grade / worst_outcome_grade computed over the non-null grades.

Long / tidy tables (the analytical heart)

• data/grades.csv — one row per compound × outcome.
• data/legal_status.csv — one row per compound × region.
• data/citations.csv — the DOI provenance table, joined back to the source text.
• data/cosmetic_claims.csv — permitted and forbidden claim wordings per cosmetic ingredient.
• data/identifiers.csv — chemical identifiers with confidence and provenance.

Parquet mirrors — data/parquet/*.parquet for all of the above (Hugging Face auto-generates Croissant JSON-LD from these).

Retrieval corpus

• data/corpus.jsonl — one record per compound: {slug, name, title, url, text}, where text is a natural-language rendering of the identity, the per-outcome grades, the legal status and the caveats. Every record carries its https://vallydia.com/compound/<slug> URL so a citation resolves back to the source.

Images — images/structures/<slug>.svg (+ .png) and images/MANIFEST.csv, which carries mandatory factual alt_text for every image.

Full field-by-field documentation: DATA_DICTIONARY.md.

The grading scale

Two axes: consistency of evidence × magnitude/quality of effect.

Grade	Meaning
A	Strong, consistent human evidence
B	Good human evidence, smaller or formulation-dependent effect
C	Mixed or limited human evidence
D	Weak — mechanistic or very small studies only
F	Either untested/unproven, or tested and failed — two different meanings that share a letter
null	The row is not an efficacy claim at all (safety and skin-penetration rows, for instance). Carried as null; never coerced into a letter.

Grades are assigned per outcome, not per compound. overall_grade reflects the leading, best-supported application, graded honestly downward and never up. See METHODOLOGY.md and vallydia.com/methodology.

Chemical identifiers, and why some are blank

Identifiers are resolved against PubChem by exact name/synonym match. A wrong identifier is worse than a blank one in an evidence dataset, so the rule is: no confident, unambiguous match → blank field and identifier_confidence: none. The 21 blanks are blends, stacks, multi-component biologics and proprietary analogs that have no single structure to point at. Where an ambiguity is real, it is recorded rather than resolved by guesswork — see entity_note in identifiers.csv (GHK-Cu is the live example: PubChem's record is the free tripeptide ligand, not the copper(II) complex, and the dataset and the image alt-text both say so).

Structure images are rendered only from high-confidence SMILES. Molecules too large for an honest 2D depiction (large peptides, GLP-1-class analogs) get a factual data card instead of an unreadable hairball.

Dataset creation

Curation rationale. Vallydia maintains an evidence-graded ingredient register because the cosmetic and research-compound market is dominated by claims that outrun their evidence. Publishing the register as open data is the most direct way to say: here is our evidence assessment — audit it.

Source data. Published literature synthesized by Vallydia into the compound register at vallydia.com. The dataset is generated from that register by build/generate_dataset.py — it is not hand-written, and re-running the generator on the same source reproduces it byte-for-byte.

Annotations. The A–F grades and the cosmetic-claims maps are produced by Vallydia. reviewer_status is unassigned throughout: no external reviewer has signed off, and the dataset does not pretend otherwise.

Considerations for using the data

• Not medical advice, and not a dosing reference. Appearance and evidence context only.
• Grades are an assessment, not a measurement. They encode a judgment about a literature base; reasonable experts may grade differently. The base, effect and caveat fields exist so you can check the reasoning rather than take the letter on faith.
• Coverage bias. The register covers compounds Vallydia's audience actually asks about — cosmetic actives and popular research peptides. It is not a systematic sample of chemical space.
• Literature bias. Grades inherit the biases of the published literature: small samples, industry-funded cosmetic trials, and publication bias toward positive findings. The caveat field records these where they are known.
• Regulatory data ages. legal_status and status_note reflect the date in last_updated. Verify anything regulatory against the primary source before relying on it.

Usage

import pandas as pd

compounds = pd.read_parquet("data/parquet/compounds.parquet")
grades    = pd.read_parquet("data/parquet/grades.parquet")

# Cosmetic ingredients with at least one A-graded outcome
best = grades[grades.grade == "A"].merge(
    compounds[compounds.is_cosmetic][["slug", "name"]], on="slug"
)

from datasets import load_dataset
ds = load_dataset("vallydia/ingredient-evidence-register", "grades")

More in examples/.

License and attribution

CC-BY-4.0. You may use, remix and redistribute this data, including commercially, provided you give attribution.

Data: Vallydia Ingredient-Evidence Register (https://vallydia.com), CC-BY-4.0.

Citation

Cite the concept DOI — it always resolves to the latest version:

Bilenko, J. (2026). Vallydia Ingredient-Evidence Register (Version 1.0.0) [Data set]. Vallydia. https://doi.org/10.5281/zenodo.21364453

@dataset{vallydia_ingredient_evidence_register,
  title     = {Vallydia Ingredient-Evidence Register},
  author    = {Bilenko, Jacob},
  year      = {2026},
  publisher = {Vallydia},
  version   = {1.0.0},
  doi       = {10.5281/zenodo.21364453},
  url       = {https://vallydia.com},
  note      = {CC-BY-4.0}
}

---

Maintained by Vallydia · Methodology · Datasheet · Data dictionary