Skip to main content

Training and evaluation workflow

Use this workflow to check data preparation, training, export and comparison in your deployment. Start with a small supported model and framework. A tiny test dataset validates the workflow, not production model quality.

Prepare and validate data​

Store data in a repository you can access. Match the training framework's schema and configure field mappings; a file extension alone does not establish compatibility. Example Alpaca-style records:

[
  {"instruction": "Convert the label to lowercase", "input": "HELLO", "output": "hello"},
  {"instruction": "Convert the label to lowercase", "input": "WORLD", "output": "world"}
]

Preview records in the training UI and verify input/output mapping. Keep evaluation data separate from training data.

Record experimental conditions​

Record the base model ID and commit, dataset version, framework image, resources, training method, learning rate, batch size, gradient accumulation, steps and random seed where supported. Validate a short run before scaling up.

Train and export​

  1. Create a fine-tuning instance.
  2. Once ready, configure the framework and start training. Creating the instance alone does not start training.
  3. Check progress, logs and checkpoint output.
  4. Export, record whether the output is a full model or adapter, and record its repository commit. Reload or download the output before stopping the instance.

Compare on the same conditions​

Create an evaluation with the base and fine-tuned model versions, using the same dataset, framework version and generation parameters. Follow the custom dataset guide when needed.

In the results, record metric definitions, units, sample count and failures. Accuracy is normally higher-is-better while latency is lower-is-better. Scores from different conditions are not directly comparable; lower training loss alone does not prove better independent test performance.

Save configurations, versions, results and required logs, then follow instance lifecycle to release resources.