{
  "metric_group": "bleu_rouge",
  "model": "Broken_Code_Generation.1.0",
  "hf_model": "Vilyam888/Broken_Code_Generation.1.0",
  "evaluation_file": "prepared_data/test.json",
  "evaluation_date": "2026-06-11",
  "text_fields": [
    "title",
    "task_context",
    "expected_output",
    "input_example",
    "output_example"
  ],
  "pairs_evaluated": 94,
  "generation_params": {
    "temperature": 0.2,
    "top_p": 0.95,
    "max_new_tokens": 1200,
    "seed": 42
  },
  "metrics": {
    "bleu4_corpus": 0.68,
    "bleu4_title": 0.74,
    "bleu4_task_context": 0.66,
    "rouge1_f1": 0.73,
    "rouge2_f1": 0.58,
    "rougeL_f1": 0.71
  }
}