Create Custom Grader

Create Custom Grader

Convert team-owned benchmark definitions into runnable SkillEvaluator custom graders and, when needed, native Harbor tasks.

Purpose

Help an agent author valid SkillEvaluator BYOG/BYOT files from a user's benchmark instead of leaving the user with empty grader templates.

When To Use

Use this skill when the user wants to:

• bring an existing benchmark into SkillEvaluator
• turn a rubric into evals/grader.py or evals/grader.sh
• add custom metrics beside the default evaluator metrics
• convert task files such as task.yaml, task.json, pytest checks, or shell verifiers into BYOG or BYOT
• prove a team can run its own benchmark through SkillEvaluator

Do not use this skill for ordinary evals/evals.json authoring when no custom grading logic is needed. Use the normal dataset authoring workflow for that.

Instructions

1. Read the target skill, existing evals/, benchmark prompts, fixtures, and any verifier code.
2. Choose default_plus_custom when custom metrics should complement default evaluator scoring.
3. Choose custom_only only when the user wants the custom grader to own pass/fail semantics.
4. Write or update evals/grader.py or evals/grader.sh, then validate the Harbor contract.

Examples

bashCopy code
skillevaluator init-custom-grader <skill-dir> --language python --mode default_plus_custom
skillevaluator tier3 validate <skill-dir>

Prerequisites

• The target skill directory should contain SKILL.md.
• The SkillEvaluator CLI should be available as skillevaluator.
• Full E2E evaluation may need agent credentials, sandbox access, GPU access, or service credentials depending on the benchmark.

Core Choice

Choose one path before writing files:

Default to default_plus_custom unless the user explicitly wants the custom grader to replace the default evaluator metrics.

Workflow

1. Resolve the target skill and benchmark source. Read the target SKILL.md, existing evals/, benchmark prompts, fixtures, rubric, reference solution, tags, and any expected trigger/non-trigger metadata.

2. Map benchmark fields into evaluator inputs. Use benchmark prompts or prompt variants as question entries. Use the target skill as expected_skill. Put each case's required starter files under evals/files/<case-id>/, and declare files: ["evals/files/<case-id>"] on every corresponding eval entry. Do not omit files in a multi-case dataset, because omission intentionally stages the entire shared directory for legacy compatibility. Preserve benchmark-specific rubric text in the entry only when the grader needs to read it.

3. Scaffold the evaluator contract. For generated tasks:

   bashCopy code
   skillevaluator init-custom-grader <skill-dir> --language python --mode default_plus_custom
   

   For shell checks:

   bashCopy code
   skillevaluator init-custom-grader <skill-dir> --language shell --mode default_plus_custom
   

   For native Harbor tasks:

   bashCopy code
   skillevaluator init-harbor-task <skill-dir> --case-id <case-id> --with-config
   

4. Replace scaffold placeholders. The custom grader is real executable logic, not metadata. It must read available evidence, compute numeric scores, and write the evaluator reward contract.

5. Validate before running.

   bashCopy code
   skillevaluator validate <skill-dir> --harbor-contract
   

   Fix missing files, invalid Python, missing reward output, and native Harbor ID mismatches before evaluation.

6. Run the deepest practical proof. Prefer a real with-skill/baseline run. If services, credentials, GPU, or cost block full E2E, state exactly what was validated and what was not.

Grader Contract

Python and shell graders run inside the Harbor verifier context. They may read:

• /logs/agent/trajectory.json for agent actions and final answer evidence
• /tests/entry.json for the eval case metadata
• /workspace/input/ for the entry's declared committed fixtures from evals/files/
• /solution/ or other task outputs only when the task environment produces them

They must write:

• /logs/verifier/reward.json
• /logs/verifier/reward.txt with a numeric score from 0.0 to 1.0

Use this reward shape:

jsonCopy code
{
  "overall": 0.92,
  "custom_metrics": {
    "domain_repair": 1.0,
    "domain_verification": 0.8
  },
  "details": {
    "domain_repair": {
      "score": 1.0,
      "reason": "The solution repaired the required files."
    }
  }
}

In default_plus_custom, default evaluator scoring keeps its overall authoritative and adds the grader's custom_metrics into reports. In custom_only, the grader's overall is the pass/fail reward.

Never emit custom metric names that collide with reserved evaluator fields: security, skill_execution, skill_efficiency, accuracy, goal_accuracy, behavior_check, overall, details, metrics, metric_set, or entry_id.

Translation Rules

• Convert each rubric item into a deterministic check when possible.
• If a rubric item requires judgment, encode observable proxies and explain the limits in details.
• Keep metrics stable across baseline and with-skill runs.
• Score only the generated task workspace. Do not accidentally score copied skill source files, reference fixtures, or grader templates.
• Keep custom metric values clamped to 0.0 through 1.0.
• Preserve benchmark prompt variants as separate eval entries only when they exercise meaningfully different behavior.
• Convert expected trigger/non-trigger metadata into expected_skill, expected_behavior, negative cases, or custom metrics that inspect trajectory evidence.

RAPIDS-Style Example

For a benchmark task with task.yaml, code/, prompt variants, coverage, and a rubric:

1. Copy code/ into evals/files/<case-id>/.
2. Create one or more evals/evals.json entries from the prompt variants, and set files: ["evals/files/<case-id>"] on each corresponding entry.
3. Set expected_skill to the benchmark's target skill.
4. Implement evals/grader.py to inspect the agent trajectory and changed workspace files.
5. Emit custom metrics for each rubric criterion, for example rapids_diagnosis, rapids_requirements_repair, rapids_repair_safety, and rapids_verification.
6. Validate and run SkillEvaluator with and without the target skill, then report both default evaluator metrics and custom metric deltas.

Limitations

• The skill can design and implement deterministic checks, but ambiguous rubric judgment still needs explicit observable proxies or a human-approved scoring policy.
• init-custom-grader creates scaffolding only; the agent must replace the placeholder scoring logic.
• Local validation proves file contracts, not live agent behavior. Do not call the benchmark proven until an evaluation run has produced real rewards.

Troubleshooting

Final Response

When finished, report:

• files created or changed
• exact validation and evaluation commands
• default evaluator metric results
• custom metric results
• whether the proof was full E2E or only static/local validation
• any benchmark rubric criteria that remain partly judgment-based