IRIS8 Talk to us

Platform · 01 Train

A model trained on your own documents

We pick a licensed open-weight model and fine-tune it on examples of your team's work, inside your boundary.

From examples to model

From your examples to your model

Four things happen, in this order.

Inside your boundary

Your documents

Past cases
Reviewed outputs
Policies

Candidates

Candidate AOpen-weight, licence checked
Candidate BOpen-weight, licence checked
Candidate COpen-weight, licence checked

Result

YoursYour model

What happens

What we do, and with what

  • Choose the base model: licensed, open-weight, sized to your hardware
  • Prepare the data with your subject experts
  • Fine-tune with low-rank adapters, then preference tuning on your reviewers' ratings
  • Version and score every run
IRIS8 console · loan-file-qa / trainingSample data

Training run · loan-file-qa

Base model
Candidate B, open-weight, licence checked
Training data
12,400 examples, personal data masked
Method
Low-rank adapters, then preference tuning
Run
v3, 14 Oct 2026
Status
Scored. Waiting for benchmark sign-off

Your decisions

You decideIRIS8 does
Which task the model learnsShortlists and tests the base models
Which examples count as done wellCleans, masks and versions the data
What data the model may seeRuns and scores each training run

You receive: the fine-tuned weights · a benchmark report on your cases · training records · a deployed model with an API

Your decisions, our work

Take this with you

Data readiness checklist

Ten checks to run before you pick a workflow to train on.

IRIS8 · Platform · Train1 page

Data readiness checklist

  • The task is repeated often and has a right answer a skilled person would agree on
  • You have past inputs for the task (files, forms, messages)
  • You have the outputs a skilled employee produced for those inputs
  • Someone can say which of those outputs were good
  • You know which fields are personal or sensitive
  • You know which of those fields the task does not need
  • The documents sit in systems you control
  • One subject expert can give a few hours a week to review examples
  • One person can approve access to the data
  • You know what the model must never see

iris8.ai/platform/train.html

Questions

Questions

Talk to us
Q1How much data do we need?

It depends on the task. We size it in the first week, and well-chosen examples matter more than volume.

Q2Fine-tuning or retrieval?

Often both. Fine-tuning teaches a fixed format, your terms and your rules; retrieval supplies facts that change, such as policies and prices.

Q3Does our data leave our environment during training?

No. In your data centre or your cloud account, training runs inside it. For cloud we manage, see the Deploy page.

Q4Which base models do you use?

Licensed open-weight models. We shortlist several for the workflow and test them on your benchmark first.

Q5Who owns the result?

You own the fine-tuned weights, the datasets and the training records.

Next step

Bring one task and a folder of examples