Evaluation Resource Card for <resource-id>
Template only. This file contains no benchmark examples or answers.
Identity
- Task:
<task-definition> - Version/revision:
<immutable-id> - Canonical repository and paper:
<links> - Licence and access terms:
<reviewed-terms> - Gating or leakage restrictions:
<restrictions>
Population and scope
Define languages, domains, time range, difficulty, participant or source population, and what the resource does not measure.
Protocol
Pin prompts, tools, retrieval index, system policy, model/checkpoint, tokenizer, decoding, environment, seeds, retries, timeouts, and scoring harness.
Metrics
For every metric, define inputs, formula, aggregation, uncertainty, invalid-run handling, and the claim it can and cannot support.
Integrity and contamination
Document split separation, train/eval overlap checks, answer-access controls, test-set secrecy, and how public traces or screenshots avoid revealing restricted examples.
Limitations
Document construct validity, annotation disagreement, judge bias, saturation, temporal drift, contamination risk, subgroup gaps, and cost constraints.
Reporting checklist
- Exact revision and harness recorded
- Licence and gating terms rechecked
- No restricted example or answer exposed
- Invalid and failed runs counted
- Confidence or run-to-run variation reported where relevant
- Capability, safety, robustness, and efficiency reported separately