108 lines
3.6 KiB
JSON
108 lines
3.6 KiB
JSON
{
|
|
"count": 100,
|
|
"solved": 1,
|
|
"solved_rate": 0.01,
|
|
"solved_at_n_count": 1,
|
|
"solved_at_n_rate": 0.01,
|
|
"pass_at_1_count": 1,
|
|
"pass_at_1_rate": 0.01,
|
|
"candidate_count_total": 100,
|
|
"candidate_count_mean_per_problem": 1,
|
|
"num_samples_max": 1,
|
|
"num_samples_min": 1,
|
|
"reason_counts": {
|
|
"numbers are not used exactly once": 32,
|
|
"value is 222, expected 88": 1,
|
|
"ok": 1,
|
|
"value is 830, expected 77": 1,
|
|
"value is 17.4025974026, expected 44": 1,
|
|
"value is 272, expected 17": 1,
|
|
"value is 44.9770114943, expected 96": 1,
|
|
"value is 36, expected 20": 1,
|
|
"value is 99, expected 30": 1,
|
|
"value is 26.5, expected 76": 1,
|
|
"value is 94.08, expected 24": 1,
|
|
"value is 270, expected 90": 1,
|
|
"value is 934, expected 48": 1,
|
|
"value is 72.8, expected 100": 1,
|
|
"value is 1216, expected 17": 1,
|
|
"value is 30, expected 26": 1,
|
|
"value is 23.0322580645, expected 44": 1,
|
|
"value is 40.2906976744, expected 52": 1,
|
|
"value is 61, expected 75": 1,
|
|
"value is 495, expected 28": 1,
|
|
"value is 1664, expected 96": 1,
|
|
"value is 32, expected 36": 1,
|
|
"value is 156, expected 98": 1,
|
|
"value is 1320, expected 92": 1,
|
|
"value is 241, expected 94": 1,
|
|
"value is 63, expected 28": 1,
|
|
"value is 16.2608695652, expected 33": 1,
|
|
"value is 715, expected 78": 1,
|
|
"numbers must be positive integers": 1,
|
|
"value is 624, expected 83": 1,
|
|
"value is 642, expected 53": 1,
|
|
"value is 56, expected 22": 1,
|
|
"value is 924, expected 65": 1,
|
|
"value is 350, expected 26": 1,
|
|
"value is 225, expected 50": 1,
|
|
"value is 960, expected 80": 1,
|
|
"value is 15.5833333333, expected 38": 1,
|
|
"value is 107.666666667, expected 61": 1,
|
|
"value is 129, expected 65": 1,
|
|
"value is 308, expected 58": 1,
|
|
"value is 102, expected 37": 1,
|
|
"value is 99, expected 100": 1,
|
|
"value is 54.3396226415, expected 36": 1,
|
|
"value is 501, expected 83": 1,
|
|
"value is 259, expected 47": 1,
|
|
"value is 120, expected 32": 1,
|
|
"value is 496, expected 100": 1,
|
|
"value is 840, expected 62": 1,
|
|
"value is 310, expected 41": 1,
|
|
"value is 390, expected 43": 1,
|
|
"value is 14, expected 15": 1,
|
|
"value is 28.5185185185, expected 71": 1,
|
|
"value is 131, expected 14": 1,
|
|
"value is 42, expected 30": 1,
|
|
"value is 9.35, expected 24": 1,
|
|
"value is 52.1276595745, expected 26": 1,
|
|
"value is 88, expected 26": 1,
|
|
"value is 143, expected 35": 1,
|
|
"value is 777, expected 59": 1,
|
|
"value is 2494, expected 66": 1,
|
|
"value is 992, expected 60": 1,
|
|
"value is 440, expected 15": 1,
|
|
"value is 38.8153846154, expected 83": 1,
|
|
"value is 192.705882353, expected 53": 1,
|
|
"value is 96, expected 78": 1,
|
|
"value is 727, expected 87": 1,
|
|
"value is 32.0641025641, expected 58": 1,
|
|
"value is 12.253164557, expected 51": 1,
|
|
"value is 426, expected 92": 1
|
|
},
|
|
"empty": false,
|
|
"first_attempt_count": 100,
|
|
"first_attempt_r1_format_rate": 1.0,
|
|
"first_attempt_answer_tag_rate": 1.0,
|
|
"first_attempt_allowed_chars_rate": 1.0,
|
|
"first_attempt_parse_ok_rate": 0.99,
|
|
"first_attempt_numbers_ok_rate": 0.67,
|
|
"first_attempt_target_ok_rate": 0.01,
|
|
"first_attempt_valid_rate": 0.01,
|
|
"attempt_count": 100,
|
|
"attempt_r1_format_rate": 1.0,
|
|
"attempt_answer_tag_rate": 1.0,
|
|
"attempt_allowed_chars_rate": 1.0,
|
|
"attempt_parse_ok_rate": 0.99,
|
|
"attempt_numbers_ok_rate": 0.67,
|
|
"attempt_target_ok_rate": 0.01,
|
|
"attempt_valid_rate": 0.01,
|
|
"mean_candidates_until_solution": 1,
|
|
"median_candidates_until_solution": 1,
|
|
"max_candidates_until_solution": 1,
|
|
"solved_at_1_count": 1,
|
|
"solved_at_1": 0.01,
|
|
"num_samples": 1
|
|
}
|