/eval
Reference for the /eval command: invocation, arguments, inputs and outputs.
Runs one evaluation campaign for a language model or prompt: scores each output over a dataset and reports metrics by category plus regressions against a baseline. Use when the user asks to 'run an eval campaign', 'evaluate this prompt' or 'check for regressions'. A reusable harness is eval-harness.
Invocation
/eval [--dataset PATH] [--scorer NAME]Pipeline position
Diagnostic surface for the model-quality review track. It consumes the evaluation dataset, the scorer definition, and the candidate endpoint, and emits the campaign findings artifact that release-readiness consumers read. The command mutates neither the candidate nor the dataset; the findings are read-only diagnostics.
Inputs
| Argument | Type | Required | Description |
|---|---|---|---|
--dataset PATH | Path | No | Path to the evaluation dataset (JSONL / CSV / Parquet). MUST carry an input column, an expected column where the scorer is reference-based, and a category column for the per-category breakdown. When omitted, Phase 1 surfaces the dataset choice as an inquiry. |
--scorer NAME | Enum | No | Scorer to apply — exact-match · semantic-similarity · rubric-graded · pairwise-preference · operator-defined. When omitted, Phase 1 surfaces the scorer choice as an inquiry with the candidate set annotated per rules/option-annotation.md. |
Next step
Invoke /perf-audit to measure the candidate endpoint's runtime against the per-class budgets once the model-quality campaign is clean; the regression ledger at _inputs/eval-findings.md is the prerequisite evidence the release-readiness sign-off consumes.
Source
Generated from src/apothem/commands/eval.md, the command definition every harness installs.