Getting started with alallo

Create an alallo account

Open the alallo dashboard and click "Sign up free" in the top bar. The person who signs up becomes the owner (the "root" user) of a new alallo account. The account owner can later invite colleagues as additional users. If you already have an alallo login, click "Sign in" instead. See Accounts and users for the roles in an alallo account.

Look around before signing up

You can use the alallo dashboard without an account to see the available models, the training methods, their settings, and time and price estimates for public Hugging Face datasets. When you are not signed in, the "Start training" button reads "Sign up to start training", and choosing a file to upload takes you to sign-up.

Add tokens to your alallo balance

alallo jobs are paid for with alallo tokens from your account's prepaid token balance. Before you start your first job, the account owner (root user) tops up the balance by card, on a secure Stripe payment page. See Tokens and pricing for how tokens work and how to top up. [TBD: where the top-up button is in the dashboard and whether new accounts get any free starting tokens.]

Open the New job form

In the alallo dashboard, click "New job" in the top bar. The New job form has these sections: Algorithm (the training method), Model, Data, parameters for the chosen method, and Goal (goal and data residency). The "Jobs" tab lists every job in your account.

Step 1: choose a training method in alallo

In the Algorithm section of the New job form, pick one training method card:

• Supervised fine-tuning, with a Precision dropdown: "Full precision (16-bit)" (LoRA) or "4-bit" (QLoRA). Use this to teach the model from example texts or instruction/answer pairs.
• Supervised fine-tuning (full weights): updates every weight of the model. Highest quality ceiling; costs the most.
• Direct preference optimisation (LoRA): teaches the model to prefer one answer over another, from pairs of a chosen and a rejected answer to the same prompt.

If you are unsure, start with Supervised fine-tuning at full precision (16-bit). See Models and training methods for plain-language explanations of each method.

Step 2: choose a base model in alallo

In the Model section of the New job form, give the job a name and choose the base model from the "Base model" list. Smaller models train faster and use fewer tokens; larger models can learn more complex behaviour. The list only shows models that alallo can train right now for the data residency you have chosen.

Step 3: choose your training data in alallo

In the Data section of the New job form, choose "Hugging Face dataset" and type a public dataset name (for example tatsu-lab/alpaca), or choose "Upload a file (JSONL, JSON or CSV)" and upload a file or pick one you uploaded before. Then:

• For supervised fine-tuning, write a text template that turns each row into one training text, using {column} placeholders.
• For DPO, enter the names of the prompt, chosen and rejected columns.
• Optionally set the split, for example train or train[:1000] for the first 1000 rows.

See Training data formats for details and examples.

Step 4: review the training settings in alallo

The parameters section shows the settings of the chosen method, such as Max steps, Learning rate and LoRA rank, already filled with sensible defaults. You can start with the defaults. "Max steps" is the setting that most affects the price and the training time. The "Require bf16" checkbox under the method cards is on by default. See Models and training methods for what each setting does.

Step 5: choose goal and data residency in alallo

In the Goal section of the New job form:

• Goal: "Lowest cost" (finish for the least money; usually also quick for small jobs) or "Fastest finish" (finish as soon as possible; the price can be a little higher).
• Data residency: "European Union", "United States" or "Anywhere (lowest price)". See Data residency and privacy.

There is no budget to set: a job can start when your available token balance covers its quoted price, and you never pay more than the quoted price.

Step 6: get the estimate and start the alallo job

Click "Estimate time and price". alallo reads and tokenizes your data and shows the job's price in alallo tokens (fixed) and estimated time (a range, such as "40 min to 1.3 h"). If capacity is busy, the estimate says "No capacity is free right now; the job would wait in the queue."

Click "Start training…" and confirm with "Yes, start". alallo reserves the quoted tokens from your balance. If you cancel the job early you pay only for the training done, plus the 1M token start fee per job; if the job fails, you pay nothing. If your balance is too low, alallo says "This job needs N tokens and your balance has M. Top up to start it." Changing any field after the estimate clears the estimate, so the job you start always matches the price you saw.

Step 7: follow the alallo job

After you start, the dashboard opens the Jobs tab with your new job selected. The Jobs tab shows each job's status, progress, estimated time left and price, and the selected job's training log. See Job status and time estimates. When the job finishes, the dashboard shows "Training finished. The fine-tuned model is ready." See Job outputs.