Skip to main content
Evaluation

Evaluation tasks

Create evaluation tasks on QwenCloud to quantify model output quality using custom dimensions and datasets.

Evaluation tasks apply evaluation dimensions to specific data and models, producing quantified results. You can create and manage tasks on the Evaluation Tasks console page.

Create an evaluation task

1

Open the create page

On the evaluation tasks list page, click Create. The system auto-generates a task name (format: eval_YYYYMMDD_HHMMSS), which you can modify.
2

Select the data source

Choose a data source type from the dropdown. The two sources differ in upload entry point and subsequent configuration:
Data sourceUpload entryDescriptionData format (JSONL)Use cases
Evaluation datasetCreate and publish on the Datasets pageThe system runs inference with the evaluation model, then scores with evaluation dimensionsPrompt (question) + Completion (reference answer)First-time evaluation, comparing multiple models
Inference resultsUpload on the Evaluation pageScores directly with evaluation dimensions, skipping inferencePrompt + Output (model output), Completion (reference answer) optionalAlready have inference results, reducing inference cost
3

Select data and configure

The configuration options shown vary by data source:
  • Evaluation dataset: Choose a published evaluation dataset from the dropdown, then select the evaluation model to run inference on the dataset. Click Manage datasets to create or publish a dataset.
  • Inference results: Upload or select existing inference results. No evaluation model is needed (the results already contain model output) — proceed directly to selecting evaluation dimensions.
Recommended data volume: 50-100 samples for initial validation, 200-500 for reliable statistical results, and 500+ for comprehensive evaluation covering edge cases.
4

Select evaluation dimensions

Choose one or more evaluation dimensions. Click Manage dimensions to go to the dimensions management page.
5

Review billing details

Expand Billing details to view estimated costs. The scoring process is free; inference is billed at standard token rates.
6

Submit the task

Click Create to submit the evaluation task.

Manage evaluation tasks

The evaluation task list displays the following information:
ColumnDescription
Task nameName of the evaluation task
ScoreScores per evaluation dimension (numeric for numeric type, pass rate for classification)
StatusTask execution status
Evaluation modelThe model used for inference
DatasetThe associated evaluation dataset
DimensionsEvaluation dimensions used
TokensToken consumption during inference
ActionsView details, delete, etc.

View evaluation results

Click Details in the actions column to open the results page, which has two tabs:
  • Metrics — Aggregated performance per evaluation dimension. Numeric dimensions show average score, score distribution, and pass rate; classification dimensions show label proportions and overall pass rate. Useful for quickly comparing scores across dimensions.
  • Data details — Per-sample scoring breakdown, including the input prompt, model output (uploaded output for inference results, online inference result for evaluation datasets), reference answer, and each dimension's score along with the judge model's reasoning. Filter by dimension or by pass/fail to locate samples where the model underperformed.

Next steps