IRIS8 Talk to us
Solutions

Fine-tune an LLM on your company data, in your own environment

We fine-tune an open-weight model on examples of how your team does a task, inside your cloud or data centre, and benchmark it against the model you use now.

This is one step of four. See how training fits the whole service: Train.

Book a call

Fine-tuning or RAG?

What data you need

  • Input/output pairs showing the task done well. OpenAI's guide suggests starting with about 50 well-made examples and adding more if results improve (OpenAI, Supervised fine-tuning guide); LIMA showed 1,000 carefully chosen examples can go a long way (Zhou et al., "LIMA").
  • Quality matters more than volume. Data prep is the biggest share of the work — data scientists report spending about 45% of their time on loading and cleaning data (Anaconda, 2020 State of Data Science). We do this work with your subject experts.
  • Personal data you don't need is masked before training.

How we fine-tune

  1. Benchmark first: we build a test set from your real cases and agree a pass mark.
  2. Choose a base model: licensed open-weights models, sized to your hardware.
  3. Train efficiently: LoRA adapters (Hu et al., "LoRA"), which train a small fraction of parameters, then preference tuning with DPO (Rafailov et al., "DPO") using your reviewers' ratings.
  4. Distil where it helps: a larger model can teach a smaller one; distilling step-by-step matched larger models with less data (Hsieh et al., "Distilling Step-by-Step").
  5. Score and compare: every run versioned and scored against your current model and frontier models.
  6. Deploy and reinforce: in your data centre, your cloud account, or cloud we manage in your region; retrained on reviewed feedback.

Public results on narrow tasks: across 310 fine-tuned models and 31 tasks, fine-tuned small models averaged 10 points above GPT-4 (Zhao et al., "LoRA Land"). In our work, specialised models perform on par with frontier models on narrow, domain-specific tasks.

Book a call

What you receive

  • The fine-tuned weights, owned by you.
  • A benchmark report on your cases.
  • Training records: data versions, runs, evals, sign-offs.
  • A deployed model with an API your apps can call.

Start with one task

Book a call

Related: 8-week pilot · Private LLM for enterprise · On-prem and VPC deployment · How to train an LLM on your own data

Frequently asked questions

Can we fine-tune an LLM on our company data without it leaving our network?

Yes. Training runs in your cloud account or data center. That is the main difference from fine-tuning through a public API.

How much data do we need?

Often hundreds to a few thousand good examples for one task (OpenAI, Supervised fine-tuning guide; Zhou et al., LIMA). We check what you have in week 1.

Isn't OpenAI fine-tuning simpler?

It was. OpenAI has closed fine-tuning to new users and allows training only until 6 Jan 2027 (OpenAI, Supervised fine-tuning guide). A model you own can't be withdrawn.

Will fine-tuning add new knowledge?

It's better at teaching behaviour than facts. For facts that change, add retrieval.

How long does it take?

One workflow goes from scoping to production in our 8-week pilot.

Is this a private LLM?

Yes, see private LLM for enterprise.

Start with one workflow.

A 30-minute call with the IRIS8 team about your data, your environment and a first pilot.

Book a call