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.

What changes
  • 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
What doesn't change
  • 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).

Step 01

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.

Step 02

Training

LoRA/QLoRA on your infrastructure or hosted GPU.

Step 03

Evaluation

Side-by-side comparison against the base model. We measure what changed.

Step 04

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?

Base model

“Yes, you can share it using secure file transfer protocols. Make sure to encrypt the data and use a secure channel.”

Fine-tuned

“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-tuningKnowledge (RAG)
What it changesHow the model behavesWhat the model knows
Best forTone, format, style, refusal patternsFacts, documents, current information
Data neededCurated examples of desired behaviorOrganized documents
UpdatesRetrain when behavior needs to changeAdd documents anytime
They work togetherFine-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