We don't label data.We supply the experts who know French HR.
Frontier models still fail at French payroll, labour law and collective agreements, because almost no one has the experts to teach them. Talma Labs sources, tests and deploys France's HR specialists for the labs and data platforms training the next generation of models.
6,000 French HR professionals in network. Tested, not self-declared. Based in France, under EU law.
A dense, regulated domain
nobody has covered.
The frontier moved to expert data
Public data is exhausted. Labs now buy judgment: professionals who write, correct and score model outputs in their own field, and who define what "correct" means. The value of an RL environment sits in the rubric, not in the software.
French HR is the hard case
Payroll, DSN filings, URSSAF, labour law, 400+ collective agreements. Rules that change every year, vary by branch, and cannot be inferred from an English-language corpus. Generalist data vendors don't cover it — and can't fake it.
We already have the people
Talma has spent years building a network of 6,000 French HR professionals — payroll managers, employment lawyers, HRBPs, recruiters. Not a marketplace we hope to fill. A network we already run.
Four ways to work with us
Expert supply
Tested French HR experts, onboarded onto your platform and your tooling. You run the missions; we guarantee the supply and the bar. Fastest route to volume.
Preference & correction data
A/B ranking of model outputs, corrected answers, question–answer pairs — produced by practitioners, double-reviewed, with inter-annotator agreement reported per batch.
Evaluation rubrics
Grading criteria written by domain specialists for each HR sub-domain — what a correct payslip correction, a valid dismissal procedure or a compliant DSN filing actually looks like.
RL environments
Verifiable simulations of French HR work: payroll and DSN processing, collective-agreement cases, full recruitment loops. Built with the experts who do the job.
Tested, not self-declared
Anyone can claim ten years in payroll. Our experts prove it before they touch a single task.
Practical test per domain
Correct a payslip seeded with errors. Rule on a rupture conventionnelle. Rewrite a CV. Scored per sub-domain, not one global grade.
Identity & fraud checks
Verified identity and right to work before onboarding. The infiltration problems that hit this market don't reach your data through us.
Structured skill profile
Level, specialities, collective agreements mastered, sectors, reliability score, mission history. Machine-readable, so matching isn't guesswork.
Calibration & agreement
The same cases are scored by several experts to measure consistency. We report inter-annotator agreement with every delivery.
Fourteen HR sub-domains
Need depth on one of these, or something adjacent? Tell us what you're training →
French payroll has a verifiable reward
Most enterprise domains are hard to turn into RL environments because correctness is a matter of opinion. French payroll isn't. A payslip is either compliant with the applicable collective agreement and the current URSSAF rates, or it isn't. A DSN filing either validates, or it is rejected with a code.
That gives you what RL actually needs: a deterministic reward signal on a long-horizon, multi-step task, in a domain with real economic value and no public training corpus.
The hard part isn't the simulator. It's the people who know which of the 400+ collective agreements applies, and why. That's the part we've spent years assembling.
French data, French experts, EU law
Models trained on data labelled outside the EU inherit that jurisdiction — even when inference is sovereign. For high-risk systems, the EU AI Act's data governance requirements (Art. 10) reach back into the sourcing pipeline. HR is explicitly a high-risk category.
Talma Labs works with experts based in France, under EU jurisdiction, with a documented audit trail for every selection and rejection decision, GDPR-compliant handling of expert data, and no demographic profiling at any stage.
If your compliance team has to answer for where your HR training data came from, that answer should be short.
From experts to environments
Vetted expert supply
Tested French HR experts onboarded onto your platform, at the volume and cadence you set. For labs and for the data platforms already serving them.
Rubrics & preference datasets
Per-sub-domain grading criteria and scored preference pairs. Demonstration datasets: annotated payslips, rewritten CVs, labour-law cases. Sold to labs and to HR tech teams fine-tuning their own models.
The French HR environment
A verifiable simulator of the job: payroll and DSN processing with computable reward, collective-agreement reasoning, end-to-end recruitment loops.
What we're learning
Benchmarks and evaluations from inside the French HR expert market.
How well do frontier models handle a French payslip?
We ran the tests. The results are worse than you'd expect →
RatesWhat French HR experts earn on AI training work
2026 rate benchmark, by sub-domain →
RegulationEU AI Act Article 10 and HR training data
Why high-risk classification changes your sourcing →
Straight answers
Do you compete with Mercor, Surge or Handshake?
Usually not. They're horizontal platforms with their own tooling and lab relationships; we're deep on one domain they don't cover. Most of the time we're the supply layer underneath — our tested French HR experts, onboarded onto their platform. We also work with labs directly when the project is HR-specific.
Why only HR? Why only France?
Because depth beats breadth in this market, and because that's where we're genuinely unmatched. French payroll and labour law are dense, high-value and structurally under-covered by English-first data vendors. We'd rather own one domain completely than be the tenth option in ten.
How do you know your experts are actually good?
Scored practical tests per sub-domain before onboarding, identity verification, and ongoing calibration against other experts on shared cases. You get inter-annotator agreement figures with every batch, not a trust-us statement.
Can you deliver in languages other than French?
Our experts work in French and English. French HR expertise is the product; English is how we deliver documentation, rubrics and reporting.
How does pricing work?
Hourly for expert supply, per-deliverable for rubrics and datasets, scoped project for environments. Contracts, NDAs, IP assignment and expert payment are handled by Talma — you receive one invoice.
How fast can we start?
Scoping takes days. A first tested cohort typically lands within [X] days of an agreed specification, depending on how narrow the sub-domain is.
Training a model on HR work?
Tell us what you're building. We'll come back with a sourcing plan and a sample rubric, not a rate card.
You know French HR?
Payroll managers, employment lawyers, HRBPs, recruiters — paid remote work teaching the models that will handle your job. On your schedule.