Capability 03 — Fine-tuning
Your model, your data, your behavior.
An off-the-shelf model speaks generic. Fine-tuning shapes it to your organization — your terminology, your tone, your compliance requirements. We modify open-weight models efficiently using LoRA and QLoRA. One GPU. Days, not months.
Overview
What is fine-tuning?
Fine-tuning is partial retraining of an open-weight model on your data. The model keeps its general intelligence — language, reasoning, knowledge — but learns to behave differently.
- Tone — Formal for legal, conversational for customer service
- Terminology — Your industry jargon, product names, internal acronyms
- Output format — Structured JSON, specific templates, compliance language
- Behavior — Refusal boundaries, escalation rules, domain-specific reasoning
- The model's core capabilities — language, reasoning, knowledge
- Your data stays yours — the fine-tuned model runs on your infrastructure
Method
How we do it.
LoRA / QLoRA: parameter-efficient fine-tuning. We train a small adapter (0.1–1% of the model's total parameters) while the base model stays frozen. Fast (hours to days), cheap (single GPU — QLoRA reduces VRAM needs by ~4×), portable (adapter files are megabytes, easy to version and A/B test), and stackable (multiple adapters per base model).
Data preparation
We help you curate and format training examples. Quality over quantity — a thousand well-crafted examples typically beat a hundred thousand noisy ones.
Training
LoRA/QLoRA on your infrastructure or hosted GPU.
Evaluation
Side-by-side comparison against the base model. We measure what changed.
Deployment
The fine-tuned model replaces or runs alongside the base model.
Example
Before and after.
Can I share this patient data with a third-party lab?
“Yes, you can share it using secure file transfer protocols. Make sure to encrypt the data and use a secure channel.”
“Sharing patient data with third parties requires a signed BAA and HIPAA-compliant transfer. Under our data governance policy (section 4.2), all external data sharing must be approved by the compliance officer. Would you like me to retrieve that policy?”
The fine-tuned model knows your rules, speaks your language, and refuses appropriately. The base model gives generic — and potentially risky — advice.
Applications
Use cases by industry.
Banking & Finance
Fine-tune for compliance tone. The model learns to refuse appropriately, cite regulations, and format outputs for audit.
Healthcare
Fine-tune for medical terminology in Spanish (or English). Diagnostic assistance, clinical documentation, patient communication.
Utilities & Energy
Fine-tune for operational protocols. The model understands grid topology, outage procedures, and regulatory reporting formats.
Legal
Fine-tune for jurisdiction-specific language. Puerto Rico law, US federal, LATAM regulatory frameworks.
Customer Service
Fine-tune for your brand voice and escalation rules. When to answer, when to hand off to a human.
Comparison
Fine-tuning vs. Knowledge (RAG) — when to use which.
| Fine-tuning | Knowledge (RAG) | |
|---|---|---|
| What it changes | How the model behaves | What the model knows |
| Best for | Tone, format, style, refusal patterns | Facts, documents, current information |
| Data needed | Curated examples of desired behavior | Organized documents |
| Updates | Retrain when behavior needs to change | Add documents anytime |
| They work together | Fine-tune for behavior… | …RAG for knowledge. |
Contact
What behavior do you want your AI to learn?
Bring us your use case. We'll tell you if fine-tuning is the right tool — and what data you'll need.
Discuss your fine-tuning needs →