Upload stage3/bird/t5-base/run_1/spider_forgetting_check/eval_metrics.json with huggingface_hub
Browse files
stage3/bird/t5-base/run_1/spider_forgetting_check/eval_metrics.json
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "t5-base",
|
| 3 |
+
"dataset": "spider",
|
| 4 |
+
"stage": "stage3",
|
| 5 |
+
"run": 1,
|
| 6 |
+
"checkpoint": "stage3/bird/t5-base/run_1/epoch-12-em0_0495",
|
| 7 |
+
"eval_beams": 4,
|
| 8 |
+
"n_samples": 1034,
|
| 9 |
+
"em_raw": 0.1277,
|
| 10 |
+
"em_norm": 0.1489,
|
| 11 |
+
"ex": 0.3859,
|
| 12 |
+
"ex_errors": 370,
|
| 13 |
+
"valid_sql_rate": 0.6422,
|
| 14 |
+
"table_precision": 0.8375,
|
| 15 |
+
"table_recall": 0.8904,
|
| 16 |
+
"table_f1": 0.843,
|
| 17 |
+
"by_sql_type": {
|
| 18 |
+
"AGGREGATION": {
|
| 19 |
+
"total": 551,
|
| 20 |
+
"em_norm": 0.0708,
|
| 21 |
+
"ex": 0.3739
|
| 22 |
+
},
|
| 23 |
+
"ORDER_BY": {
|
| 24 |
+
"total": 237,
|
| 25 |
+
"em_norm": 0.1899,
|
| 26 |
+
"ex": 0.4937
|
| 27 |
+
},
|
| 28 |
+
"LIMIT": {
|
| 29 |
+
"total": 189,
|
| 30 |
+
"em_norm": 0.0952,
|
| 31 |
+
"ex": 0.4392
|
| 32 |
+
},
|
| 33 |
+
"SIMPLE": {
|
| 34 |
+
"total": 185,
|
| 35 |
+
"em_norm": 0.3189,
|
| 36 |
+
"ex": 0.3676
|
| 37 |
+
},
|
| 38 |
+
"GROUP_BY": {
|
| 39 |
+
"total": 277,
|
| 40 |
+
"em_norm": 0.0108,
|
| 41 |
+
"ex": 0.3357
|
| 42 |
+
},
|
| 43 |
+
"NESTED": {
|
| 44 |
+
"total": 83,
|
| 45 |
+
"em_norm": 0.1928,
|
| 46 |
+
"ex": 0.253
|
| 47 |
+
},
|
| 48 |
+
"JOIN": {
|
| 49 |
+
"total": 408,
|
| 50 |
+
"em_norm": 0.0,
|
| 51 |
+
"ex": 0.2868
|
| 52 |
+
},
|
| 53 |
+
"HAVING": {
|
| 54 |
+
"total": 79,
|
| 55 |
+
"em_norm": 0.0,
|
| 56 |
+
"ex": 0.3165
|
| 57 |
+
}
|
| 58 |
+
},
|
| 59 |
+
"by_hardness": {
|
| 60 |
+
"easy": {
|
| 61 |
+
"total": 456,
|
| 62 |
+
"em_norm": 0.2961,
|
| 63 |
+
"ex": 0.4693
|
| 64 |
+
},
|
| 65 |
+
"medium": {
|
| 66 |
+
"total": 402,
|
| 67 |
+
"em_norm": 0.0075,
|
| 68 |
+
"ex": 0.3433
|
| 69 |
+
},
|
| 70 |
+
"hard": {
|
| 71 |
+
"total": 168,
|
| 72 |
+
"em_norm": 0.0952,
|
| 73 |
+
"ex": 0.2798
|
| 74 |
+
},
|
| 75 |
+
"extra_hard": {
|
| 76 |
+
"total": 8,
|
| 77 |
+
"em_norm": 0.0,
|
| 78 |
+
"ex": 0.0
|
| 79 |
+
}
|
| 80 |
+
},
|
| 81 |
+
"error_types": {
|
| 82 |
+
"WRONG_COLUMNS": 183,
|
| 83 |
+
"EXTRA_JOIN": 170,
|
| 84 |
+
"OTHER": 161,
|
| 85 |
+
"WRONG_TABLE": 312,
|
| 86 |
+
"MISSING_JOIN": 32,
|
| 87 |
+
"MISSING_GROUP_BY": 18,
|
| 88 |
+
"MISSING_WHERE": 2,
|
| 89 |
+
"MISSING_ORDER_BY": 2
|
| 90 |
+
}
|
| 91 |
+
}
|