{
  "selection": "best_loss",
  "metric": "eval_loss",
  "maximize": false,
  "value": 0.08017373085021973,
  "step": 392,
  "epoch": 1.0051397365884998,
  "training_seed": 2026,
  "split_seed": 2026,
  "adapter_dir": "qwen35-9b-grading-checkpoint-comparison/run-20261009T143514-201079Z/seed_2026/best_loss",
  "validation_metrics": {
    "eval_loss": 0.08017373085021973,
    "eval_grading_examples": 269,
    "eval_grading_valid_examples": 269,
    "eval_grading_recovered_output_rate": 0.0,
    "eval_grading_invalid_output_rate": 0.0,
    "eval_grading_exact_match_accuracy": 0.7174721189591078,
    "eval_grading_per_criterion_accuracy": 0.8262548262548263,
    "eval_grading_precision": 0.8563049853372434,
    "eval_grading_recall": 0.8768768768768769,
    "eval_grading_f1": 0.8664688427299704,
    "eval_grading_final_score_exact_accuracy": 0.758364312267658,
    "eval_grading_final_score_within_one_accuracy": 0.9888475836431226,
    "eval_grading_final_score_mae": 0.2527881040892193,
    "eval_grading_final_score_rmse": 0.5244930283645093,
    "eval_grading_final_score_pearson": 0.7658867012264351,
    "eval_grading_final_score_qwk": 0.7626225285601851,
    "eval_grading_question_count": 9,
    "eval_grading_question_macro_exact_match_accuracy": 0.7057324291998621,
    "eval_grading_question_macro_per_criterion_accuracy": 0.8368088075303911,
    "eval_grading_question_macro_criterion_f1": 0.8543485592790837
  }
}