Minimal Reference Dataset and Baseline #10

Objective

Provide a small, deterministic dataset that allows CI validation and regression testing.

Scope

Create a lightweight dataset suitable for pull request evaluation.

Tasks

• Define dataset format JSONL recommended

• Add 10 to 20 examples including:

  • Grounded responses
  • Hallucination traps
  • Optional tool-call examples

• Document dataset structure

• Define baseline thresholds per metric

• Implement regression detection via thresholds

Acceptance Criteria

• Dataset runs in under a few minutes
• Evaluation can fail based on threshold regression
• Dataset format is documented

Out of Scope

• Large production datasets
• Automatic dataset generation