Ground truth, from the field.
Frontenac Labs collects, labels and evaluates the data that AI models learn from. Every record carries its source, its annotator and its review history, and it stays in Canada.
Data a model team can build on, and defend.
We work for teams that train, fine-tune or test AI models and need to know where every example came from.
Datasets, sourced and labelled
Licensed text, speech and images, labelled against your guideline and delivered with a datasheet.
Tests your model has never seen
Held-out test sets and scoring rubrics built for your task, kept separate from anything used in training.
Judgement from people who know the field
Ratings, comparisons and corrections of model output from vetted reviewers, including lawyers, clinicians and engineers.
Three stages, one chain of custody.
Each project moves through the same pipeline. Nothing reaches a training run without passing through all three.
Sourced with consent
We source and license text, speech, images and sensor data with written consent and a recorded chain of custody for every file.
Labelled against a guideline
Trained annotators work from a written guideline. A second reviewer checks a sample of every batch, and disagreements go back into the guideline.
Measured on your task
We build held-out test sets and scoring rubrics, so you can measure a model on the work you care about before and after training.
Six kinds of work, one record format.
Every item, whatever its type, ships with the same provenance fields.
Text and documents
Contracts, support conversations, forms and reports, cleaned and labelled at the span, sentence or document level.
Speech and audio
Recordings and transcripts with speaker, timing and accent labels, including Canadian French and regional English.
Images and video
Boxes, masks, keypoints and frame-level tags, with a second reviewer on every batch.
Preference and ranking
Side-by-side comparisons and ratings of model answers, for fine-tuning and reward models.
Red-teaming
Prompts written to find where a model fails, logged with the failure and the expected answer.
Bilingual sets
Matched English and French data, written and reviewed by native speakers rather than machine translated.
What a delivery looks like.
Every dataset ships with a datasheet and one record per item. The figures below are illustrative.
Canadian French support conversations
- Items
- 48,000
- Languages
- fr-CA, en-CA
- Annotators
- 22
- Second review
- 10% of every batch
- Agreement (Cohen's κ)
- 0.86
- Format
- JSONL + datasheet
{
"id": "fl-ds-0142-000001",
"text": "Bonjour, ma facture de mars est en double.",
"lang": "fr-CA",
"label": "billing.duplicate_charge",
"annotator": "ann-017",
"reviewed_by": "rev-004",
"source": { "licence": "client-consented", "collected": "2026-03-11" },
"guideline": "v3.2"
}
Built around Canadian privacy law.
How we store, process and account for the data you send us.
Canadian residency TBC
Data is processed and stored on Canadian infrastructure by default. Any exception is written into your contract.
PIPEDA and Quebec Law 25 TBC
Consent records, retention schedules and access requests are handled under federal and Quebec privacy law.
Consent on file
We keep written consent and a licence for every source file, and we can show you either on request.
Provenance on every row
Each record links to its source, its licence, the person who labelled it and every review it passed.
English and French
Guidelines, annotators and test sets work in both official languages, Canadian French included.
Agreement, measured
We report inter-annotator agreement per label, so you can see where a dataset is solid and where it is thin.
Who you'll work with.
A small team of annotation leads, linguists and evaluation researchers.
Your company’s records may be worth licensing.
AI companies license years of real business records, such as documents, project files and chat history, to train their models. Our two-minute check estimates how much data you hold and whether you fit what buyers look for.
Free, no files shared. Frontenac Labs is an independent referrer and may receive a fee from the buyer if your company licenses data.
AI training data and data licensing, answered.
What does Frontenac Labs do?
We collect, label and evaluate the data AI models learn from: training datasets, evaluation sets and expert review, in English and French. Every record we deliver keeps its source, its annotator and its review history.
Can my company license its data to AI companies?
Often, yes. AI companies license years of real business records, such as documents, project files, chat history and CRM records, to train their models. Buyers usually look for 30 or more employees, a base in the United States, Canada or Europe, mostly English records and several years of history. Consumer, health, legal, payment-card and children’s data is usually not accepted. Our free Data Check gives you a rough estimate in about two minutes.
How much is company data worth to AI buyers?
It varies widely. Published deals for business data have ranged from about $100,000 to over $2 million per company. That is market context, not a quote: any offer comes from the buyer after they review your data, and some companies are not accepted.
Does Frontenac Labs get paid when a company licenses its data?
For data licensing, Frontenac Labs is an independent referrer and does not act for any buyer. If your company licenses data to a buyer we introduce, that buyer may pay us a referral fee. You pay us nothing, and we cannot share any part of that fee with you.
Can you build a bilingual French and English dataset?
Tell us the task, the languages and the volume through the form below and we will reply with a scope, including what we can and cannot deliver.
What happens to the details I enter in the Data Check?
Nothing is sent to us until you press “Request review”. To pre-fill the form, our server reads your company’s public website and suggests details, which you can correct. See our privacy policy.
Tell us what your model needs to learn.
Share the basics and we will reply with a scoping call and a sample guideline.