View evaluation results

Describes how to view evaluation task status and result details: Task Overview, Case Analysis (per-run sampling and aggregate-by-question Pass@K), result bucketing and export backflow, and how to interpret baseline evaluation reports.

View evaluation results overview

After an evaluation task is submitted, you can view the task status and evaluation results in the evaluation task list. The task list displays the task name, evaluation model, evaluation method, status, overall score, evaluation details, and creation time; the Actions column supports terminate, download, and delete; searching by name, filtering by status, and sorting by time are supported. When the task status changes to Completed, click the task name to enter the details page and view the detailed evaluation results—custom evaluation reports contain two tabs, Task Overview and Case Analysis (see below), and baseline evaluation reports contain four tabs: Task Overview, Benchmark Score Details, Case Analysis, and Multi-task Comparison (see Baseline evaluation report).

Task status

The status of an evaluation task transitions in the following order: Pending → In Progress → Completed/Failed/Terminated.

Task Overview and Case Analysis

When the task status changes to Completed, click the task name to enter the details page, which contains the Task Overview and Case Analysis tabs.

Task Overview tab: Displays the task basic information (evaluation model, inference configuration, evaluation method, evaluation dimensions, concurrency, inference times per sample, creation and update time), the overall metrics table (overall score and number of bad cases of each evaluation dimension), and the consumption summary table (model name, type, and input/output usage). Click View Details next to the inference configuration to view the System Prompt and inference parameters.

Case Analysis tab: Displays evaluation details sample by sample. The toolbar supports:

• Switching the input column display (Prompt&Completion, Prompt, or Completion) and the list view.
• Per-run sampling view: one row per inference, showing the input columns (for prompt&completion datasets, the Prompt and Completion columns; for ChatML datasets, the Messages column showing the complete multi-turn conversation), model output (Output), scores of each evaluation dimension, bucket result, and status. Click Details to view the complete multi-turn input and the scoring basis (judge model prompt and output) in a drawer.
• Aggregate by question view: aggregates the multiple inference results by sample (question), shows the Pass@K metric of each evaluation dimension (the proportion of passes across K inferences, reflecting the model stability on that sample), and you can click View Details to view the inference results of each run for the sample. This view is unavailable when the inference times per sample is 1; if any inference of a sample is unsuccessful (scoring failed, scoring terminated, or in progress), the Pass@K of that sample is displayed as -.
• Switching between Normal mode and Quick annotation for manual annotation (see Manual annotation evaluation).
• Entries for manual dimension configuration, bucketing, and export (see below).

Result bucketing and export

Bucketing classifies evaluation samples by rules, which helps locate difficult samples and optimize them in a targeted manner:

• Bucketing rules: On the Case Analysis tab, click Bucketing to configure rules (bucketing is supported only for completed evaluations). Buckets are divided by the score interval of the selected metric; an interval includes its lower bound but not its upper bound. If a sample matches multiple intervals, the first bucket from top to bottom takes effect; samples that match no bucket are assigned to Others (not matching the preceding rules). Click Add Bucket to add a bucket and select the bucketing metric (such as Pass@K). In the aggregate-by-question view, samples whose Pass@K is - do not participate in bucketing.
• Bucket results: After bucketing is applied, a Bucket Result column is added to the case list, showing the bucket of each sample.
• Export: Click the export icon to open the Export Bucket Results dialog. The export scope can be Export All or Download by Bucket; the export destination can be Download to Local (generates a data file containing buckets, scores, and sampling details, in XLSX format) or Backflow to Platform Dataset (creates a new platform dataset from the selected buckets for subsequent training or evaluation). When bucketing is not applied, only exporting all is supported.

Baseline evaluation report

After a baseline evaluation task is completed, click the task name to enter the report page, which contains four tabs: Task Overview, Benchmark Score Details, Case Analysis, and Multi-task Comparison. After the evaluation is completed, you can click "Download" in the Actions column of the task list row to download the evaluation report.

Task Overview: Displays basic information (evaluation model, inference configuration, evaluation method, creation time, update time, task progress, evaluation dataset), overall metrics (a radar chart for intuitive comparison of scores across evaluation sets), and a score table for each evaluation set (listing the evaluation dataset, completed/total samples, overall score, and number of bad cases).

Benchmark Score Details: At the top, switch by benchmark and select a drill-down dimension (dimensions vary by benchmark and are dynamically returned by the system based on the dimensions of each benchmark). It displays a score bar chart and a details table under the selected dimension (evaluation dataset, subject score, number of samples, and number of bad cases). The "View" action in the Actions column navigates to Case Analysis.

Case Analysis: Displays the input, reference answer, model output, and status (Pending/In Progress/Completed/Failed/Terminated) of each case, as well as the evaluation result (Pass/Fail). Click a specific case to view the complete input, reference answer, model output, and evaluation basis (the LLM judge's rationale).

Multi-task Comparison: Click "Add Model Comparison" to add historical evaluation tasks (only tasks of the same modality can be compared, up to 5). After adding, a multi-task radar comparison, comprehensive comparison table, and strengths/weaknesses analysis are generated to horizontally compare the performance of multiple models on each benchmark.

Result analysis recommendations

When analyzing evaluation results, focus on the score distribution rather than only the overall score. The overall score is the average score across dimensions and may mask significant differences between dimensions. For example, an overall score of 85 may be the average of 95 in some dimensions and 75 in others—only per-dimension analysis can reveal weaknesses.

Score differences of 1-3% are usually evaluation noise rather than model capability differences; decisions should not be made based on small score gaps alone. Tasks using manual evaluation dimensions change to the Completed status only after all data is annotated. For the detailed workflow, see Manual annotation evaluation.

When the inference times per sample is greater than 1, you can use Pass@K to determine sample difficulty: samples with a high pass rate across multiple inferences are content the model has stably mastered, while samples with a low pass rate are difficult cases. Difficult cases can be filtered with bucketing and backflowed as platform datasets (viewable in Data Management) to expand evaluation sets or fine-tuning data.

FAQ