[
  {
    "label": "Output v0",
    "file": "output_v0.jsonl",
    "description": "Baseline export generated from the initial evaluation run."
  },
  {
    "label": "Output v1",
    "file": "output_v1.jsonl",
    "description": "Updated question generation and validation scripts."
  },
  {
    "label": "Output v2",
    "file": "output_v2.jsonl",
    "description": "Latest export after changing/refining the plan generation."
  },
  {
    "label": "Outputs from 2.5 Flash",
    "file": "output_v3_Flash.jsonl",
    "description": "Outputs generated from the 2.5 Flash evaluation run."
  },
  {
    "label": "Sample 500 Questions",
    "file": "500_questions.jsonl",
    "description": "500 sample questions generated using gemini 2.5 Flash."
  },
  {
    "label": "Questions without FMI",
    "file": "40_questions_without_FMI.jsonl",
    "description": "40 sample questions generated when Financial Markets and Investments is not an included class."
  },
  {
    "label": "40 Questions with FMI",
    "file": "40_questions_direct.jsonl",
    "description": "40 sample questions generated when Financial Markets and Investments is an included class."
  }
  
]