Training settings
What each training setting does, how the estimate is calculated, and what happens while your model trains.
Base model
The Models page lists every model available for training. Smaller models train faster and cost less; larger models handle harder tasks. Start with the smallest model that could plausibly do the task, measure it, and move up only if you need to.
Method
Training uses supervised fine-tuning (SFT) with LoRA adapters: the model learns from your example inputs and outputs. Preference tuning (DPO) and reinforcement learning are planned and not yet available. See Tuning methods.
Settings
| Setting | Default | What it does |
|---|---|---|
| Epochs | 1 | How many times training passes through your examples. Whole numbers from 1. More epochs cost proportionally more. |
| LoRA rank | 8 | Size of the trained adapter. Higher ranks can learn more but cost more to train. |
| Learning rate | Model default | How large each training step is. |
| Batch size (samples) | Model default | Examples per training step. |
| Warmup steps | Model default | Steps over which the learning rate ramps up. |
| Context length | Model default | Longest example, in tokens, the run accepts. |
| Weight decay | Model default | Regularisation that discourages overfitting. |
| Learning rate schedule | Model default | Constant, linear or cosine. Linear and cosine also accept a minimum learning-rate ratio and a decay ratio. |
Leave a setting on its default unless you have a reason to change it. Settings a model does not support are rejected before the run starts.
Evaluation
- Automatic evaluation split (default): a small part of the training data is set aside to measure progress during training.
- Use the prepared validation split: your 15% validation split is used instead. See Splits.
Your 15% test split is never used for training or evaluation during training.
The cost estimate
The sidebar estimates the cost before you start:
training tokens = tokens in your training split × epochs
estimated cost = training tokens ÷ 1,000,000 × rate per million tokensThe rate depends on the base model's size. The sidebar shows the rate used for your estimate, which is also the rate your run is charged at.
The estimate is not the charge. You pay the verified cost of the run, which can be higher or lower than the estimate, for example when tokenisation differs from our count.
What happens when you start
- The estimate plus 25% is held from your credits, and your run is queued, then picked up within seconds.
- The training split is prepared and training starts. Progress appears under Training.
- When training finishes, the run shows Trained and billing review is pending until the final cost is verified.
While a run is in progress you cannot remove your card, and you can start another usage-billed run only after any earlier run's bill is settled.
If the run fails before training starts, it ends as failed and nothing is charged. If it fails after training started, any verified usage is still billed.