Dataset opportunity

Tolid — LLM-annotated corpora: crypto news, biomedical patents, beauty video

A data company selling the corpora it built — and still operating the pipeline that built them.

LLM-annotated corpora / knowledge graphstextDomain LLM training, knowledge graphs, sentiment models🌍 Francetolid.ioJul 15, 2026

Confidence

60%

Market

Annotated training data whose label quality is MEASURED, not asserted — a claim almost no competing corpus can make. Sold with its defects declared, and with the pipeline that produced it available as a service.

Lineage

How this lead was derived

The signal-first chain, end to end: recent external signals → qualified niche → resolved data-holder → site verification → scored opportunity. Every lead is explainable.

Profile

Dataset profile

Type

LLM-annotated corpora / knowledge graphs

Modality

text

Sector

AI training data

Volume

2.47M docs · 723k patents · 449 routines

Freshness

Archive — pipelines restartable

Rarity

High — the annotation, not the text

Accessibility

Immediate — BigQuery / Firestore mirrors

Legal

Annotations, graphs and metadata are derived works produced by Tolid and are transferable. The source press text is NOT (it belongs to the publishers) — it is excluded from every delivery. BIOPORTAL — measured and settled (2026-07-14): 707 ontologies are used, not one. The 13,757 rows (1.6% of the mapping table) coming from restrictively licensed sources — SNOMED CT, MedDRA, RxNorm, OMIM, NDDF, Read Codes, ICD, CPT — are EXCLUDED from every delivery. Everything else is open by construction (OBO Foundry mandates CC-BY/CC0 of its members; MeSH is free for commercial use). CC-BY is an ATTRIBUTION obligation, not a copyleft: an attribution NOTICE ships with the data. STILL OPEN: the Open Beauty Facts ODbL status (beauty corpus) — adversarial review refuted the 'Produced Work' reading, so the constraint may prove BINARY rather than a price discount. The four biometric inference columns (perceived sex, age range, emotion, skin tone) are excluded from any delivery.

Buyer persona

Publishers of domain LLMs and financial-NLP models · pharma and biotech R&D (patent intelligence) · cosmetics groups and beauty-tech · data vendors who want to train their own classifier. NOT hedge funds or signal vendors for the crypto corpus — it carries no alpha, and we say so.

Tolid is a data company. It does not resell someone else's exhaust: it built the collection and LLM-annotation pipelines that produced these corpora, and it still operates them. What is on offer is therefore two things at once — a stock of annotated data, and the machine that produced it.

Three corpora are on the table.

1. Crypto news, annotated (2,469,218 documents, Oct 2023 → Sep 2025). Sentiment, topic, summary and a buy/hold/sell tag, on normalised tickers. Read this before anything else: the trading signal has NO measured predictive power. We ran the backtest against real prices ourselves, on all 1,198,479 signals: it beats a coin flip by 0.7 points, and its market-neutral alpha is nil. The reason is in the data — press sentiment correlates at +0.60 with PAST returns and +0.07 with future ones. It follows the price; it does not lead it. We are not selling alpha, and this corpus is not for a trading desk. What that same +0.60 does prove is that the annotation is faithful: it is an LLM-labelled financial corpus whose label quality has been validated against an external ground truth. That is what you are buying — training data with a measured quality, not a signal.

2. Biomedical patents, structured as a knowledge graph (723,149 documents). The most valuable asset of the three, and the one carrying the largest open question: the licensing terms of the BioPortal ontologies it aligns to are under review. We say so before you ask.

3. Procedural beauty video (449 routines, 7,656 timestamped gestures, 21,441 videos). Not "video data": an analytical surface over what people actually do, step by step, with which products. Two constraints declared up front — the Open Beauty Facts ODbL question is unresolved and may be binary rather than a discount, and the four biometric inference columns (perceived sex, age, emotion, skin tone) are excluded from any delivery. The procedural graph keeps all of its value without them.

What we do not sell: the source article text. It belongs to the publishers, not to Tolid. What is transferable is the derived work — the annotations, the URLs, the metadata, the graph.

Services

What the holder can also do for you

This dataset is not only a stock — its holder still operates the pipeline that produced it. Each capability below is backed by an observed fact, not a claim.

Human validation of the annotations, on demand

Tolid can run a human review pass over any corpus it has produced — full or sampled, to your own guidelines and quality bar.

EvidenceThe entire human-in-the-loop layer already exists — nine moderation tables, an escalation queue, prompt structures — and has never been run on a single document (0 of 4,841,938). The machine is built and wired; it has simply never been switched on.

A corpus built to your theme, from scratch

Give a domain and a taxonomy; Tolid runs collection, LLM annotation and graph construction end-to-end — the same pipeline that produced these corpora.

EvidenceThe pipeline is theme-driven, and there is proof: a seventh theme ("Geopolitical Tensions and Alliances") is fully configured — prompt and JSON template written — with no collection and no deployed service. A theme conceived, written, never launched. That is what "give us a subject" looks like in this codebase.

Entity resolution and normalisation

Mapping the messy surface forms of a domain onto canonical entities — the unglamorous work that decides whether a corpus is joinable with your own data.

Evidence21,359 alias → canonical-ticker mappings were built for the crypto corpus; 19,080 semantic mappings for the beauty ontology. This is the domain's price of entry, already paid.

Cleaning and extending an existing signal

Re-parsing, de-duplicating and extending a field that was produced once and never curated.

EvidenceDeclared openly: the buy/hold/sell field covers only ~44% of the crypto rows and carries parsing leaks (`buy (implied)`, `hold/sell`). Tolid can clean and extend it. We would rather sell you the fix than hide the defect.

Scoring

Scored dimensions

Explainable, evidence-based dimensions (0–100). The radar shows the investment axes.

-

See dimension details
SpecificityRarityVolumeTraining ValueBuyer DemandEvidence StrengthData Orientation

Evidence

Dataset evidence & lineage

What the typed evidence proves the company holds — reframed for clarity and set against the market.

-

Marketplace

Dataset details

Geographic coverage

Global (multilingual sources)

Time range

2023-10 → 2025-09 (crypto) · patents and beauty: see each dataset

Delivery

Secure extract download, or scoped access

Formats

parquet, csv, jsonl

License

Derived works (annotations, graphs, metadata) are transferable. Source press text is not.

Personal data

No PII

· licence· final price on request

There is no public price grid for these corpora, and we will not manufacture one. Our own valuation engine returns wide, UNANCHORED ranges — dispersion up to ×7 on the crypto asset across three identical runs — and a dedicated deep-research study (July 2026, 103 agents, 86 claims → 25 adversarially verified) confirmed WHY: for an LLM-annotated news archive sold as annotations-only, no defensible market comparable exists. The published editor↔LLM deals disclose amounts but never volumes, so no per-document price can be derived; and the closest academic comparable (Financial PhraseBank) anchors label PEDIGREE, never a number. The gap is in the market, not in our research. So we price ON REQUEST, against two things we can defend: a measured production cost (from real cloud billing) and a label quality validated against external ground truth. The biomedical-patents asset — the strongest of the three — is under its own valuation study; its price will be shown only once anchored on real dataset transactions, never on a SaaS-subscription proxy. You do not buy the bundle: each dataset is priced on its own.

Detailed schema & sample available on access request.

Want this data?

Request access — we broker a secure deal room. Operator-reviewed, no automatic sharing.

Share this opportunity

This listing was generated automatically from public signals. It is not verified, and we are not affiliated with this company.

Deal room

Tolid — LLM-annotated corpora: crypto news, biomedical patents, beauty video

status: broker_protected

There is no public price grid for these corpora, and we will not manufacture one. Our own valuation engine returns wide, UNANCHORED ranges — dispersion up to ×7 on the crypto asset across three identical runs — and a dedicated deep-research study (July 2026, 103 agents, 86 claims → 25 adversarially verified) confirmed WHY: for an LLM-annotated news archive sold as annotations-only, no defensible market comparable exists. The published editor↔LLM deals disclose amounts but never volumes, so no per-document price can be derived; and the closest academic comparable (Financial PhraseBank) anchors label PEDIGREE, never a number. The gap is in the market, not in our research. So we price ON REQUEST, against two things we can defend: a measured production cost (from real cloud billing) and a label quality validated against external ground truth. The biomedical-patents asset — the strongest of the three — is under its own valuation study; its price will be shown only once anchored on real dataset transactions, never on a SaaS-subscription proxy. You do not buy the bundle: each dataset is priced on its own.

Coverage

Scanned sources

https://tolid.iodiscovered

Deliverable

Premium dataset report

A teaser is generated for each dataset opportunity. The full report is available on unlock.

Teaser is public · premium is locked behind access.

From the marketplace

Explore live data opportunities

Browse datasets by sector & use-case