Train on Facts
You Can Prove
Teams build vertical AI training data on top of Prometheux: rules run over your data derive verified, task-specific examples for fine-tuning. Experts review the rules once, and every example traces to the rules and records behind it.
| question | answer | check |
|---|---|---|
| Is Lyon exposed? | Yes · tier 2 | ✓ rule |
| Is Turin exposed? | Yes · tier 3 | ✓ rule |
| Is Porto exposed? | No | ✓ rule |
Proven in production
Labels Drift.
Rules Don't.
A vertical model is only as good as the examples it learns from. Hand-labelled and scraped data is slow, inconsistent and impossible to audit. Reasoning-generated data is none of those.
Expert Logic In.
Training Data Out.
Encode what your experts know as an ontology, run it over your data, and every derived fact becomes a candidate training example: a question, a verified answer and the reasoning path between them. The approach is published in our RuleML+RR paper on fine-tuning enterprise LLMs via ontological reasoning.
Derived, Not Scraped
Examples come from rules applied to your records, so they reflect your domain and nothing else.
Reasoning Paths Included
Chase provenance gives the chain of rules and facts behind each answer, ready to teach a model the why as well as the what.
Coverage by Design
Recursion reaches the deep, multi-hop cases that sampling and manual labelling tend to miss.
Scale on GPUs
Vadalog Parallel distributes reasoning across clusters and GPUs, as in AstraZeneca's 9× faster biomedical reasoning.
Build vertical models on data you can audit. Turn proprietary know-how into a model advantage. Regenerate training data when the rules change.
From Rules to a
Fine-Tuned Model
Encode the Domain
Capture entities, relationships and expert rules over your connected sources.
Reason Over the Data
Run the rules with provenance on, deriving every conclusion and the path that led to it.
Shape the Examples
Your team turns derived facts and their derivations into task-specific question, answer and rationale examples, with a short script over the results.
Fine-Tune and Grade
Train your vertical model, then grade its answers against the same rules: a deterministic benchmark, not a judgement call.
Vertical Models for
Specialist Work
Analyst Copilots
Train on derived supplier risk, dependency and compliance conclusions, with the rule path as the rationale.
Biomedical Q&A
Generate question and answer pairs from reasoning over compounds, genes and diseases.
Regulation-Aware Models
Turn encoded regulatory rules into examples that teach a model how the rules apply to real cases.
Published Method
Our paper on fine-tuning enterprise LLMs with rule-based reasoning, presented at RuleML+RR 2023.
Browse the papers →“Prometheux technology powers highly scalable and explainable reasoning over some of the world's largest knowledge graphs.”
Accelerating Neurosymbolic AI with RAPIDS and Vadalog Parallel
“The ability to query heterogeneous databases through a flexible ontology makes it possible to infer deep knowledge and automate complex tasks, delivering trustworthy insights that far exceed the capabilities of traditional systems.”
Anna FavaFormer Senior Manager, Data & AI · Dompé
Questions,
Answered
What is vertical AI training?
Fine-tuning a language model for one domain or task. It is something teams build on top of Prometheux: the rules derive examples from your own data, each with the rules and facts that make it correct, and your team shapes them into a training set.
Should I fine-tune, or keep the rules executable?
Usually both. Keep decisions that must be exact as executable rules the model can call. Fine-tune where you need the model to understand your domain's language and reasoning patterns, using data the rules have verified.
How is this different from expert labelling?
Labelling and expert review judge examples one at a time. Here experts review the rules once, the rules derive the examples, and the same rules grade the fine-tuned model. Coverage grows with your data, and the corpus regenerates when a rule changes.
Can we use it to evaluate models, not just train them?
Yes. Hold back derived examples as a private test set and score the model's answers against what the rules derive. Each expected answer comes with the derivation that makes it correct.
Does my data leave my environment?
Reasoning runs over your sources in place. You decide which derived examples are exported for training and where the model is fine-tuned.
Related Capabilities
Build a Model That
Knows Your Domain
Bring your rules and a target task. We'll show you how to build a verified training corpus from your own data.
