nebula@atheron:/fine-tuning$
// Fine-tuning
An open model, yours.
LoRA fine-tuning on private hardware.
Upload your examples, pick a base model, and we train a LoRA adapter for your task. Your data stays private by default: it runs only on Atheron house nodes or AWS GPUs.
// What you get
A small adapter, a better model
LoRA trains a small set of extra weights beside the base model, so training is quicker and the result is easy to keep and move.
Your adapter
Download the trained adapter and use it with the base model wherever you run it. It is yours.
A held-out score
We keep part of your examples aside and report how the adapter does on them, so you can see it learned.
Open base models
Start from the open models in our catalogue, each listed with its licence and size.
Use it in batches
Run your adapter on Nebula batch inference without moving it anywhere.
// Your data
Private by default
Training data is often the most sensitive thing a team has. We treat it that way.
Our hardware only
Fine-tuning runs on Atheron house nodes or AWS GPUs. It never reaches a member's device unless you mark the data shareable.
Deleted on schedule
Your examples are deleted when the job ends and its retention window passes. You choose the window.
Never used to train anything else
We use your data for your job and nothing else.
Encrypted
In transit and at rest, with access only through short-lived signed links.
// How a job runs
From examples to adapter
Prepare your examples once. We handle the GPUs.
- 1
Upload examples
A file of inputs and the outputs you want, in the format the docs show.
- 2
Pick a base model
Choose from the catalogue, and set how many passes over the data to make.
- 3
Train
The job runs on our private GPUs. You can follow its progress and its loss as it trains.
- 4
Download
Get the adapter and its held-out score, or run it straight away in a batch.
Questions
- What does shareable mean?
- Data you mark shareable may run on member devices, which costs less. Only mark data shareable if it is already public or holds nothing private. Everything else stays on our hardware.
- How am I billed?
- By the GPU time the job uses, taken from your prepaid credit in Canadian dollars. The estimate is reserved before the job starts, and you pay what it uses, never more.
- Can I bring my own base model?
- Not yet. Jobs start from the models in our catalogue.
- Can I run my own container?
- Customer containers are planned for verified businesses, with a sandbox that has no network and no access to anyone's files. They are not open yet.
Tune a model for your task
When Nebula opens, you create your team in the developer portal, pass the business check, add a card, and start a job with your examples.