Training¶
Adapt an open model to your task and create versions you can deploy and evaluate.
Supervised fine-tuning (SFT) teaches a model from inputs and desired responses. It suits classification, extraction, response formatting, and domain-specific language when good examples are available. The current console's creation flow focuses on SFT / LoRA.
What a training run needs¶
| Input | How to choose |
|---|---|
| Base model | Choose an available model according to capability and cost |
| Dataset version | Use an uploaded, validated SFT dataset with valid samples |
| Configuration | Start with the UI defaults; adjust one factor at a time after reviewing results |
SFT and RL in the pricing table identify service rate categories. Training methods available to your account are determined by the current console options.
From configuration to a model version¶
The console organizes configuration into base model, dataset, and optional settings, with a summary of the proposed job. Run the training precheck and resolve blocking issues before submitting.
Serverless jobs wait for platform-scheduled resources. My training (我的训练) shows job status; job details provide curves, logs, and checkpoints. After a successful run, manage its version in My models (我的模型).
Keep the base model, dataset version, and configuration associated with each experiment so changes remain interpretable.
Decide whether training helped¶
Training curves describe optimization, not business quality on their own. Compare the base model and trained version on held-out examples with consistent scoring and inference settings. Inspect failures as well as aggregate results.
If behavior is still inconsistent, first review example quality and coverage before changing hyperparameters.
Continue from an existing version¶
Use Continue training (续训) in My models. The new run inherits the parent version's base model and uses your chosen dataset. Resuming requires a recoverable checkpoint; see Checkpoints for the distinction between inference weights and training state.