{"audit_file": "generation_audit_v3.json", "counts": {"accepted": 1223, "execution_errors": 228, "generated": 12000, "rejected": 10549}, "dataset_file": "gen_tasks_v3.jsonl", "evaluation_file": "gsm8k_test_eval.jsonl", "label_error_count": 0, "overlap_count": 0, "resolver": "Qwen/Qwen3.5-397B-A17B, thinking disabled, greedy, max_tokens 4096, blind to label and code", "sample_denominator": 100, "sample_seed": 0, "stage": "before_training", "version": 1}