data/labeled/labeling_statistics.json

{
  "total_samples": 9996,
  "label_distribution": {
    "gemini-pro": 210,
    "gpt-3.5-turbo": 7137,
    "mixtral-8x7b": 246,
    "claude-3-sonnet": 446,
    "gpt-4-turbo": 1541,
    "claude-3-opus": 416
  },
  "average_confidence": 0.5814911678956589,
  "validation_passed": false
}