Fine-tune an open model on your task
In early access: upload a few hundred examples, and call your tuned model through the same OpenAI-compatible API.
From examples to your own model
You bring the examples. We train the adapter and serve it behind the API you already use.
A JSONL file in chat format, one example per line: the input and the answer you’d have written. Start with a few hundred.
{"messages": [
{"role": "system", "content": "Summarise rulings in our house structure."},
{"role": "user", "content": "<the ruling>"},
{"role": "assistant", "content": "<the summary your team wrote>"}
]}One example per line in your file, shown wrapped here
GLM-5.3 Flash or Qwen3.8 27B. Keep the default LoRA rank, epochs and learning rate, or set your own.
- GLM-5.3 FlashZ.ai · 1M context
- Qwen3.8 27BQwen · 256K context
- LoRA rank
- default
- Epochs
- default
- Learning rate
- default
The base model stays frozen. Only a small LoRA adapter is trained, on GPUs in European data centres.
Base model, frozen LoRA adapter, trained
Change the model name in your OpenAI-compatible client. That’s the only code change.
from openai import OpenAI client = OpenAI( base_url="https://api.lyceum.technology/openai/v1", api_key=os.environ["LYCEUM_API_KEY"],) client.chat.completions.create( model="<your tuned model>", messages=[...],)Only the model name changes
Two base models to start
Both are open-weight models we already serve, so your tuned model runs on the same stack.
Made for narrow tasks
Fine-tuning suits narrow tasks with a clear right answer, where your team already has good examples.
Summarise rulings from one field of law
Rulings and the summaries your team wrote
Summaries of new rulings in your house structure
Write posts in your brand voice
Your best posts, each with the brief behind it
Posts from a brief, in each platform’s length and tone
Reply in your support team’s tone
Resolved tickets with the reply your agents sent
Draft replies that follow your policies on refunds and delays
Extract fields from documents
Claims, invoices or contracts with the fields you need
Your schema for a new document, with your rules for edge cases
Questions before you start
Can I use it today?
Fine-tuning is in early access. Book a call and we’ll set up your first runs with you.
How many examples do I need?
Start with a few hundred good ones. On the call we look at your set with you and say whether it’s enough.
What does it cost?
We agree pricing with the first customers. On the call we go through your use case and the numbers.
Where does my data go?
Your dataset is stored in European data centres and used only for your training jobs. How long we keep it, and who on our side can access it, we set with you in the early-access agreement.
Do I own the tuned model?
The adapter it produces is yours: a small file of LoRA weights you can download and run elsewhere.
Can I tune a model that isn’t on the list?
Not yet. Tell us which model you need on the call: that’s what early access is for.