Upload bird/bart-base/run_1/eval_metrics.json with huggingface_hub
Browse files
bird/bart-base/run_1/eval_metrics.json
CHANGED
|
@@ -1,62 +1,67 @@
|
|
| 1 |
{
|
| 2 |
-
"model": "
|
| 3 |
"dataset": "bird",
|
|
|
|
| 4 |
"run": 1,
|
| 5 |
-
"checkpoint": "bird/bart-base/run_1/epoch-
|
| 6 |
"eval_beams": 4,
|
| 7 |
"n_samples": 1534,
|
| 8 |
-
"em_raw": 0.
|
| 9 |
-
"em_norm": 0.
|
| 10 |
-
"ex": 0.
|
| 11 |
-
"ex_errors":
|
| 12 |
"ex_missing_db": 0,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
"by_difficulty": {
|
| 14 |
"simple": {
|
| 15 |
"total": 925,
|
| 16 |
-
"em_norm": 0.
|
| 17 |
-
"ex": 0.
|
| 18 |
},
|
| 19 |
"moderate": {
|
| 20 |
"total": 464,
|
| 21 |
-
"em_norm": 0.
|
| 22 |
-
"ex": 0.
|
| 23 |
},
|
| 24 |
"challenging": {
|
| 25 |
"total": 145,
|
| 26 |
"em_norm": 0.0,
|
| 27 |
-
"ex": 0.
|
| 28 |
}
|
| 29 |
},
|
| 30 |
"by_sql_type": {
|
| 31 |
"ORDER_BY": {
|
| 32 |
"total": 305,
|
| 33 |
-
"em_norm": 0.
|
| 34 |
-
"ex": 0.
|
| 35 |
},
|
| 36 |
"LIMIT": {
|
| 37 |
"total": 293,
|
| 38 |
-
"em_norm": 0.
|
| 39 |
-
"ex": 0.
|
| 40 |
},
|
| 41 |
"JOIN": {
|
| 42 |
"total": 1140,
|
| 43 |
-
"em_norm": 0.
|
| 44 |
-
"ex": 0.
|
| 45 |
},
|
| 46 |
"AGGREGATION": {
|
| 47 |
"total": 659,
|
| 48 |
-
"em_norm": 0.
|
| 49 |
-
"ex": 0.
|
| 50 |
},
|
| 51 |
"NESTED": {
|
| 52 |
"total": 118,
|
| 53 |
-
"em_norm": 0.
|
| 54 |
-
"ex": 0.
|
| 55 |
},
|
| 56 |
"GROUP_BY": {
|
| 57 |
"total": 120,
|
| 58 |
-
"em_norm": 0.
|
| 59 |
-
"ex": 0.
|
| 60 |
},
|
| 61 |
"HAVING": {
|
| 62 |
"total": 21,
|
|
@@ -65,17 +70,17 @@
|
|
| 65 |
},
|
| 66 |
"SIMPLE": {
|
| 67 |
"total": 119,
|
| 68 |
-
"em_norm": 0.
|
| 69 |
-
"ex": 0.
|
| 70 |
}
|
| 71 |
},
|
| 72 |
"error_types": {
|
| 73 |
-
"WRONG_TABLE":
|
| 74 |
-
"OTHER":
|
| 75 |
-
"MISSING_JOIN":
|
| 76 |
-
"
|
| 77 |
-
"
|
| 78 |
-
"
|
| 79 |
-
"
|
| 80 |
}
|
| 81 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"model": "bart-base",
|
| 3 |
"dataset": "bird",
|
| 4 |
+
"stage": "stage1",
|
| 5 |
"run": 1,
|
| 6 |
+
"checkpoint": "bird/bart-base/run_1/epoch-09-em0_0228",
|
| 7 |
"eval_beams": 4,
|
| 8 |
"n_samples": 1534,
|
| 9 |
+
"em_raw": 0.0339,
|
| 10 |
+
"em_norm": 0.0346,
|
| 11 |
+
"ex": 0.0737,
|
| 12 |
+
"ex_errors": 1252,
|
| 13 |
"ex_missing_db": 0,
|
| 14 |
+
"valid_sql_rate": 0.1838,
|
| 15 |
+
"table_precision": 0.7831,
|
| 16 |
+
"table_recall": 0.7337,
|
| 17 |
+
"table_f1": 0.7362,
|
| 18 |
"by_difficulty": {
|
| 19 |
"simple": {
|
| 20 |
"total": 925,
|
| 21 |
+
"em_norm": 0.0541,
|
| 22 |
+
"ex": 0.1092
|
| 23 |
},
|
| 24 |
"moderate": {
|
| 25 |
"total": 464,
|
| 26 |
+
"em_norm": 0.0065,
|
| 27 |
+
"ex": 0.0194
|
| 28 |
},
|
| 29 |
"challenging": {
|
| 30 |
"total": 145,
|
| 31 |
"em_norm": 0.0,
|
| 32 |
+
"ex": 0.0207
|
| 33 |
}
|
| 34 |
},
|
| 35 |
"by_sql_type": {
|
| 36 |
"ORDER_BY": {
|
| 37 |
"total": 305,
|
| 38 |
+
"em_norm": 0.0426,
|
| 39 |
+
"ex": 0.0623
|
| 40 |
},
|
| 41 |
"LIMIT": {
|
| 42 |
"total": 293,
|
| 43 |
+
"em_norm": 0.0444,
|
| 44 |
+
"ex": 0.0648
|
| 45 |
},
|
| 46 |
"JOIN": {
|
| 47 |
"total": 1140,
|
| 48 |
+
"em_norm": 0.014,
|
| 49 |
+
"ex": 0.0386
|
| 50 |
},
|
| 51 |
"AGGREGATION": {
|
| 52 |
"total": 659,
|
| 53 |
+
"em_norm": 0.0288,
|
| 54 |
+
"ex": 0.1032
|
| 55 |
},
|
| 56 |
"NESTED": {
|
| 57 |
"total": 118,
|
| 58 |
+
"em_norm": 0.0169,
|
| 59 |
+
"ex": 0.0508
|
| 60 |
},
|
| 61 |
"GROUP_BY": {
|
| 62 |
"total": 120,
|
| 63 |
+
"em_norm": 0.0333,
|
| 64 |
+
"ex": 0.0667
|
| 65 |
},
|
| 66 |
"HAVING": {
|
| 67 |
"total": 21,
|
|
|
|
| 70 |
},
|
| 71 |
"SIMPLE": {
|
| 72 |
"total": 119,
|
| 73 |
+
"em_norm": 0.1261,
|
| 74 |
+
"ex": 0.1345
|
| 75 |
}
|
| 76 |
},
|
| 77 |
"error_types": {
|
| 78 |
+
"WRONG_TABLE": 701,
|
| 79 |
+
"OTHER": 513,
|
| 80 |
+
"MISSING_JOIN": 97,
|
| 81 |
+
"MISSING_WHERE": 24,
|
| 82 |
+
"MISSING_ORDER_BY": 22,
|
| 83 |
+
"EXTRA_JOIN": 107,
|
| 84 |
+
"MISSING_GROUP_BY": 17
|
| 85 |
}
|
| 86 |
}
|