Adapt any language model to your domain in hours, without a GPU and without machine learning skills.
The Digital Partner is a tool for digital sovereignty. In local mode, no data leaves your infrastructure. The QLoRA service is an outsourced adaptation service — distinct from the local Digital Partner. To stay true to our confidentiality commitment, we offer two modes:
You run the fine-tuning on your own machine, with your own data. Included from the Essential version.
If you don't have a GPU or the required resources, the provider marketplace connects you with qualified experts. Digital Partner only steps in as backup if no provider is available.
Important distinction: what Digital Partner does locally (memory, governance, proofs) is different from what the QLoRA service does (GPU fine-tuning). The first is 100% sovereign — no data leaves. The second is an outsourced adaptation service, offered as an option for those without a GPU, and requires explicit consent to grant access to their data.
QLoRA is a semantic shortcut. In most business cases, it is the right choice — it offers the best quality/cost ratio without requiring an expensive GPU. But depending on your context, RAG or classic LoRA may fit better. Answer the 4 questions below for an informed recommendation.
Normally detected when you install Digital Partner.
QLoRA excels on semantic data (text, FAQ, documents). Classic LoRA is preferable if you need to teach functions or highly structured tasks.
Collect between 100 and 2,000 Q&A pairs representative of your business, in JSON or CSV format. That's all you need to get started.
[
{"question": "What is the retention period for employment contracts?",
"answer": "5 years after the end of the contract, except for special cases."},
{"question": "How is severance pay calculated?",
"answer": "1/4 of monthly salary per year of tenure up to 10 years, then 1/3 beyond."}
]Submit your dataset through the web interface (SaaS) or hand it to a partner provider (Marketplace). The QLoRA job runs in 1 to 4 hours on our dedicated GPU infrastructure.
A lightweight weight file (a few hundred MB) is stored on our servers and immediately accessible through the Digital Partner API. You can also export it for offline use (Extended offer).
The general model knows the law, but not your firm. With QLoRA, it learns your jargon, your deadlines, your procedures, and answers in clear language adapted to your clients.
Thousands of questions about your products? The model learns your products, your FAQs, your brand tone — and becomes your automated support in your voice.
Your technical reports are highly specific. The model learns your vocabulary, your domain's priorities, and your deliverable style — accessible summaries guaranteed.
Self-sufficient or need an expert? Digital Partner offers both routes, on the same sovereign infrastructure.
Describe your need, qualified providers respond. You pick the one that fits. The Marketplace is a SaaS service available to Essential and Extended subscribers.
Safety net — bootstrapping: if you find no provider within 5 business days, Digital Partner steps in directly. This safety net will fade progressively as the Marketplace grows.
Why no displayed price? Every project is unique. Providers will quote you according to your real need.
Reserved for subscribers (Essential or Extended).
The Digital Partner is an infrastructural trusted third party:
What is QLoRA?
QLoRA (Quantized Low-Rank Adaptation) is a technique that adapts a language model to a specific domain by modifying a small fraction of its parameters. It is light, efficient, and does not require retraining the whole model.
How long does adaptation take?
Between 1 and 4 hours depending on dataset size and allocated GPU power.
Is my data secure?
Yes. Submitted data is kept only for the duration of training and deleted after adaptation. We never use it for any other purpose.
Can I use the adapted model elsewhere?
Yes, the adapted model can be exported and used offline (Extended offer).
Why choose Digital Partner?
14-day free trial. No commitment.
Start now →