Csandal17/lagomy-uk-supplements

Lagomy UK Supplement Dataset (v0.1)

A structured, dose-level dataset of UK food supplements, transcribed by hand from manufacturer labels and brand-published nutrition tables.

50 products, growing. This is a working sample, not a finished corpus. It exists because there is no adequate open dataset of UK supplement compositions: the available options are US-regime databases, retailer scrapes, or aggregator sites, none of which are reliable for UK products.

---

Provenance methodology

This is the part that matters, and it is the reason this dataset exists at all. Every design decision below was made to keep the data traceable back to a specific label.

The governing principle: show, link, source, never conclude

The dataset records what a label says. It does not record what that means, whether a dose is adequate, whether two ingredients interact, or whether a product is worth taking. There are no verdicts, scores, rankings, or derived safety judgements anywhere in these files, and there will not be in later versions. Where a downstream user wants an interpretation, the job of this dataset is to hand them the sourced text and let them make it.

This is a deliberate constraint, not an unfinished feature.

Source hierarchy

Data is captured from the most authoritative available source, in this order, and the tier used is recorded per product in data_source:

1. label_photo — a photograph of the physical pack, including peel-back panels. This is the only tier that guarantees the text is contiguous as printed. The photograph is often taken in-store, which does not weaken it: what matters is that the source is the printed panel itself rather than anyone's transcription of it.
2. brand_website — the manufacturer's own published nutrition table.
3. retailer — a major retailer's listing, used only where the brand does not publish the product at all.

In this release: 34 products at brand_website, 16 at label_photo, none at retailer.

Some products exist in no published source at all. Several of the pack-photographed entries here are products whose manufacturers publish no nutrition table on their own site, and which are not listed with full composition anywhere online. They cannot be scraped, from any source, by anyone. They are in this dataset because someone stood in a shop and photographed the panel.

Brand site beats retailer listing. Two independent failure modes drove this rule. Retailers keep stock on shelf until expiry, so a listing can describe an older generation than the current formulation; and separately, a retailer's website copy can be stale while its stock is current. One product in this set (P041) was caught by the second case: the retailer listing described a formulation the pack contradicted.

A brand page can also contradict itself. Where a brand's ingredient prose and its nutrition table disagree, the table is preferred. Tables are regulated content and appear to be maintained more carefully than marketing prose. Two products here (P040, P049) show this pattern. Where the table is also the only amount-bearing source, prose-only items are not captured at all rather than being recorded without quantities.

Ingredient identity

The canonical ingredient name is the chemical form named in the ingredients paragraph, not the everyday nutrient name from the nutrition table. So ferrous fumarate, not iron; zinc oxide, not zinc. The everyday name is kept in synonyms.

A different form is always a new dictionary entry, never a merge. Ferrous fumarate and iron bisglycinate are different substances with different absorption profiles, and collapsing them to "iron" would destroy the main thing this dataset is for. The same rule separates an extract from a whole powder (ashwagandha root extract vs ashwagandha), a specified vitamin K form from an unspecified one, and food-state forms from isolated ones.

Where they exist, entries carry CAS numbers and UNII identifiers, which makes that rule mechanical rather than a matter of judgement. Metal salts are keyed to the anhydrous parent unless a label states the hydrate.

Identifiers are deliberately blank for substances that genuinely do not have a single one: fixed oils, multi-component botanical extracts, whole plant material, proteins and collagens, natural mixtures, and the entire food-grown block. In those rows notes says why.

Actives only

Excipients, bulking agents, carriers and capsule shells are not captured. The test is the label, not the substance: an oil is an active only when the nutrition table gives it a row and an amount. Olive oil is an active in one product here (the label credits it with 28mg omega-9) and a skipped softgel carrier in another. Same substance, different treatment, decided by the label each time.

Doses

Amount, unit and basis are stored separately (amount, unit, per) rather than as a single string. The per basis is read from the nutrition table's own header, which is the most common transcription trap in this domain: a label reading "Av. per 2 tablets" records per serving, with the serving size in the product's notes. Filing a per-serving figure as a per-tablet figure would silently double every dose in the product.

pct_nrv is left blank where no NRV exists for that nutrient. It is never invented, estimated, or carried over from a US %DV. UK/EFSA NRVs only.

Warnings and claims are verbatim

warning_text and claim_text are transcribed word for word from the label, including typos, which are preserved and flagged in notes rather than silently corrected.

Two rules protect that:

• Contiguity. Retailer mirrors reorder and concatenate label sentences, so a warning built from a retailer listing can look verbatim while joining sentences that are not adjacent on the pack. Pack photos are what make a row genuinely verbatim.
• Warning blocks are never shared between products, even near-identical siblings. Two products in this set are almost the same tub, and one carries an iron caution the other does not. Copying a block across products would fabricate label text.

Health claims are split one row per nutrient where a printed claim covers several. Inherently joint authorised claims (the EPA+DHA heart claim) stay as one row. A single nutrient with three functions also stays as one row: the rule is per nutrient, not per function.

Claims are keyed to the Great Britain nutrition and health claims register. GB diverged from the EU register in January 2021.

What is excluded

• Discontinued and superseded product generations.
• Products whose ingredients cannot be sourced from any tier. These remain in the private working sheet as barcode-only stubs and are excluded from the export until sourced, rather than being published with gaps.

The exported product list is derived from the composition table: a product appears here only if it has real ingredient rows. That is a join, not a manual flag, so unfinished entries cannot leak into a release.

---

Files

File	Rows	What it is
products.csv	50	One row per product, with brand, pack, price and provenance
ingredients.csv	149	Ingredient dictionary, with CAS and UNII where they exist
product_ingredients.csv	952	Dose-level composition. The core table
product_warnings.csv	183	Label warnings, verbatim, typed
product_claims.csv	43	Authorised health claims, verbatim
products_full.jsonl	50	The same data, nested one product per line

The CSVs are the source of truth. products_full.jsonl is a convenience view built from them, for anyone who would rather not do the joins.

Keys

product_id (P001…), ingredient_id (I001…) and brand_id (B001…) are stable across releases. They are the join keys between tables and will not be renumbered.

Schema

products.csv product_id, brand_id, brand_name, product_name, barcode_ean, category, form, pack_size, price_gbp, source_url, data_source, barcode_source, date_verified, notes

ingredients.csv ingredient_id, canonical_name, synonyms (semicolon separated), type (vitamin | mineral | amino acid | fatty acid | botanical | other), cas_number, unii, notes

product_ingredients.csv product_id, ingredient_id, ingredient_name, amount, unit (mg | µg), per (per tablet | per capsule | per serving), pct_nrv, notes

product_warnings.csv warning_id, product_id, warning_type (allergen | interaction | upper_limit | population | contains_caution | storage | other), warning_text, notes

product_claims.csv claim_id, product_id, ingredient, claim_text, source

A single label sentence may be dual-tagged as two warning types where it carries two hazards, so that each is independently queryable. A fish caution appears as both allergen and interaction.

Blank means absent. Empty cells are genuine nulls, not placeholders.

---

Loading it

import pandas as pd

products = pd.read_csv("products.csv")
comps = pd.read_csv("product_ingredients.csv")
ingredients = pd.read_csv("ingredients.csv")

# every product containing a specific iron form
iron = comps[comps.ingredient_id == "I005"].merge(products, on="product_id")

Or the nested view:

import json

with open("products_full.jsonl") as f:
    records = [json.loads(line) for line in f]

On the Hub:

from datasets import load_dataset

ds = load_dataset("Csandal17/lagomy-uk-supplements", "product_ingredients")

---

Known limitations

• 50 products is a sample. Coverage is weighted towards brands that publish full nutrition tables, which biases the set towards larger and more transparent UK manufacturers. Brands that hide panels under peel-off labels are under-represented, though pack photography is closing that gap.
• No formulation versioning yet. Each product is a single current record. Reformulations and pack generations are not modelled, though the build has already met real cases (a documented reformulation that cut a daily dose from three capsules to two, and a repackaging that changed label wording without changing the formula). Version-validity fields are the main planned schema change.
• Canonical dose normalisation is incomplete. A µg-based canonical amount column is planned but not populated, so cross-product dose comparison currently requires handling mg and µg yourself.
• date_verified records when the entry was checked, not when the pack was manufactured. A product verified in August 2026 may have been reformulated since.
• One product has no barcode. It is recorded as blank rather than guessed. Barcodes are never inferred: marketplace and pharmacy listings carry conflicting EANs for the same product, so an unverified barcode is left empty.
• Not a medical or regulatory resource. See below.

Intended use

Built for research, tooling and analysis: ingredient prevalence, dose distributions, label-claim patterns, formulation comparison, and as a reference layer for applications that need to identify UK supplement products.

Not suitable for clinical, diagnostic or regulatory use. Nothing here is dietary advice. Warning and claim text is transcribed from labels for reference and is not a substitute for reading the pack you actually hold, which may be a different generation from the one recorded. Verify against the physical product before relying on any figure.

Licence

CC BY-NC 4.0. Free to use, share and adapt for non-commercial purposes with attribution. For commercial licensing, get in touch.

Citation

@misc{lagomy_uk_supplements_2026,
  title  = {Lagomy UK Supplement Dataset},
  author = {Sandal, Chantal},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Csandal17/lagomy-uk-supplements},
  note   = {Version 0.1}
}

About

Lagomy is a supplement logging app in development. This dataset is its reference layer: the product and ingredient data the app is built on, published separately because the data is useful on its own.

Built and maintained by Chantal Sandal. GitHub · Hugging Face

Corrections are welcome. If a figure here disagrees with the pack in your hand, the pack is probably right and I would like to know. Open a thread in the Community tab.