Jeremiah Zhou
commited on
Commit
•
dacfb22
1
Parent(s):
d0db5a5
End of training
Browse files- all_results.json +38 -0
- eval_results.json +18 -0
- runs/Jun27_16-33-59_pikachu/events.out.tfevents.1656326997.pikachu.3190917.2 +3 -0
- test_results.json +18 -0
- train_results.json +8 -0
- trainer_state.json +1900 -0
all_results.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 3.0,
|
| 3 |
+
"eval_avg_sts": 0.7499060859363368,
|
| 4 |
+
"eval_avg_test": 0.7465428552421288,
|
| 5 |
+
"eval_avg_train_and_dev": 0.7475216864388343,
|
| 6 |
+
"eval_runtime": 20.0869,
|
| 7 |
+
"eval_sickrelatedness_spearman": 0.7077244255701862,
|
| 8 |
+
"eval_sickrelatedness_test_spearman": 0.7026718095093082,
|
| 9 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7128746659624327,
|
| 10 |
+
"eval_sts12_spearman": 0.6428011959930444,
|
| 11 |
+
"eval_sts13_spearman": 0.8168437631289807,
|
| 12 |
+
"eval_sts14_spearman": 0.7226103243605893,
|
| 13 |
+
"eval_sts15_spearman": 0.8024592797345551,
|
| 14 |
+
"eval_sts16_spearman": 0.778377683273261,
|
| 15 |
+
"eval_stsbenchmark_spearman": 0.778525929493741,
|
| 16 |
+
"eval_stsbenchmark_test_spearman": 0.7600359306951637,
|
| 17 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7821687069152359,
|
| 18 |
+
"test_avg_sts": 0.7499060859363368,
|
| 19 |
+
"test_avg_test": 0.7465428552421288,
|
| 20 |
+
"test_avg_train_and_dev": 0.7475216864388343,
|
| 21 |
+
"test_runtime": 20.1381,
|
| 22 |
+
"test_sickrelatedness_spearman": 0.7077244255701862,
|
| 23 |
+
"test_sickrelatedness_test_spearman": 0.7026718095093082,
|
| 24 |
+
"test_sickrelatedness_train_and_dev_spearman": 0.7128746659624327,
|
| 25 |
+
"test_sts12_spearman": 0.6428011959930444,
|
| 26 |
+
"test_sts13_spearman": 0.8168437631289807,
|
| 27 |
+
"test_sts14_spearman": 0.7226103243605893,
|
| 28 |
+
"test_sts15_spearman": 0.8024592797345551,
|
| 29 |
+
"test_sts16_spearman": 0.778377683273261,
|
| 30 |
+
"test_stsbenchmark_spearman": 0.778525929493741,
|
| 31 |
+
"test_stsbenchmark_test_spearman": 0.7600359306951637,
|
| 32 |
+
"test_stsbenchmark_train_and_dev_spearman": 0.7821687069152359,
|
| 33 |
+
"train_loss": 0.00036472132327159247,
|
| 34 |
+
"train_runtime": 8039.9662,
|
| 35 |
+
"train_samples": 1000000,
|
| 36 |
+
"train_samples_per_second": 373.136,
|
| 37 |
+
"train_steps_per_second": 5.83
|
| 38 |
+
}
|
eval_results.json
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 3.0,
|
| 3 |
+
"eval_avg_sts": 0.7499060859363368,
|
| 4 |
+
"eval_avg_test": 0.7465428552421288,
|
| 5 |
+
"eval_avg_train_and_dev": 0.7475216864388343,
|
| 6 |
+
"eval_runtime": 20.0869,
|
| 7 |
+
"eval_sickrelatedness_spearman": 0.7077244255701862,
|
| 8 |
+
"eval_sickrelatedness_test_spearman": 0.7026718095093082,
|
| 9 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7128746659624327,
|
| 10 |
+
"eval_sts12_spearman": 0.6428011959930444,
|
| 11 |
+
"eval_sts13_spearman": 0.8168437631289807,
|
| 12 |
+
"eval_sts14_spearman": 0.7226103243605893,
|
| 13 |
+
"eval_sts15_spearman": 0.8024592797345551,
|
| 14 |
+
"eval_sts16_spearman": 0.778377683273261,
|
| 15 |
+
"eval_stsbenchmark_spearman": 0.778525929493741,
|
| 16 |
+
"eval_stsbenchmark_test_spearman": 0.7600359306951637,
|
| 17 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7821687069152359
|
| 18 |
+
}
|
runs/Jun27_16-33-59_pikachu/events.out.tfevents.1656326997.pikachu.3190917.2
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:562ae5b39d2c0ea6f2635ed15653a0ee04e95d39273b8447cb20490d31b83a94
|
| 3 |
+
size 2056
|
test_results.json
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 3.0,
|
| 3 |
+
"test_avg_sts": 0.7499060859363368,
|
| 4 |
+
"test_avg_test": 0.7465428552421288,
|
| 5 |
+
"test_avg_train_and_dev": 0.7475216864388343,
|
| 6 |
+
"test_runtime": 20.1381,
|
| 7 |
+
"test_sickrelatedness_spearman": 0.7077244255701862,
|
| 8 |
+
"test_sickrelatedness_test_spearman": 0.7026718095093082,
|
| 9 |
+
"test_sickrelatedness_train_and_dev_spearman": 0.7128746659624327,
|
| 10 |
+
"test_sts12_spearman": 0.6428011959930444,
|
| 11 |
+
"test_sts13_spearman": 0.8168437631289807,
|
| 12 |
+
"test_sts14_spearman": 0.7226103243605893,
|
| 13 |
+
"test_sts15_spearman": 0.8024592797345551,
|
| 14 |
+
"test_sts16_spearman": 0.778377683273261,
|
| 15 |
+
"test_stsbenchmark_spearman": 0.778525929493741,
|
| 16 |
+
"test_stsbenchmark_test_spearman": 0.7600359306951637,
|
| 17 |
+
"test_stsbenchmark_train_and_dev_spearman": 0.7821687069152359
|
| 18 |
+
}
|
train_results.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epoch": 3.0,
|
| 3 |
+
"train_loss": 0.00036472132327159247,
|
| 4 |
+
"train_runtime": 8039.9662,
|
| 5 |
+
"train_samples": 1000000,
|
| 6 |
+
"train_samples_per_second": 373.136,
|
| 7 |
+
"train_steps_per_second": 5.83
|
| 8 |
+
}
|
trainer_state.json
ADDED
|
@@ -0,0 +1,1900 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_metric": 0.7475216864388343,
|
| 3 |
+
"best_model_checkpoint": "checkpoint/reproduce/unsup-bert-base-uncased/checkpoint-20000",
|
| 4 |
+
"epoch": 3.0,
|
| 5 |
+
"global_step": 46875,
|
| 6 |
+
"is_hyper_param_search": false,
|
| 7 |
+
"is_local_process_zero": true,
|
| 8 |
+
"is_world_process_zero": true,
|
| 9 |
+
"log_history": [
|
| 10 |
+
{
|
| 11 |
+
"epoch": 0.04,
|
| 12 |
+
"learning_rate": 2.2182723071453966e-06,
|
| 13 |
+
"loss": 0.0182,
|
| 14 |
+
"step": 625
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"epoch": 0.04,
|
| 18 |
+
"eval_avg_sts": 0.5761493170311456,
|
| 19 |
+
"eval_avg_test": 0.5719267013539271,
|
| 20 |
+
"eval_avg_train_and_dev": 0.5955653275065955,
|
| 21 |
+
"eval_runtime": 20.9393,
|
| 22 |
+
"eval_sickrelatedness_spearman": 0.614035610432408,
|
| 23 |
+
"eval_sickrelatedness_test_spearman": 0.6110088391931596,
|
| 24 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.6170011765653474,
|
| 25 |
+
"eval_sts12_spearman": 0.3883511001773007,
|
| 26 |
+
"eval_sts13_spearman": 0.6409334821384267,
|
| 27 |
+
"eval_sts14_spearman": 0.5233435331267315,
|
| 28 |
+
"eval_sts15_spearman": 0.6304876176600429,
|
| 29 |
+
"eval_sts16_spearman": 0.6668361049295604,
|
| 30 |
+
"eval_stsbenchmark_spearman": 0.569057770753549,
|
| 31 |
+
"eval_stsbenchmark_test_spearman": 0.5425262322522679,
|
| 32 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.5741294784478437,
|
| 33 |
+
"step": 625
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"epoch": 0.08,
|
| 37 |
+
"learning_rate": 4.440099537859936e-06,
|
| 38 |
+
"loss": 0.0009,
|
| 39 |
+
"step": 1250
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"epoch": 0.08,
|
| 43 |
+
"eval_avg_sts": 0.6176613225678802,
|
| 44 |
+
"eval_avg_test": 0.6146526994706545,
|
| 45 |
+
"eval_avg_train_and_dev": 0.6354918146843727,
|
| 46 |
+
"eval_runtime": 22.0658,
|
| 47 |
+
"eval_sickrelatedness_spearman": 0.6481782560461822,
|
| 48 |
+
"eval_sickrelatedness_test_spearman": 0.6454675815413363,
|
| 49 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.6510278427067593,
|
| 50 |
+
"eval_sts12_spearman": 0.4492677971681109,
|
| 51 |
+
"eval_sts13_spearman": 0.6749925740490339,
|
| 52 |
+
"eval_sts14_spearman": 0.5645147644908317,
|
| 53 |
+
"eval_sts15_spearman": 0.6736146568642941,
|
| 54 |
+
"eval_sts16_spearman": 0.6965736684860125,
|
| 55 |
+
"eval_stsbenchmark_spearman": 0.6164875408706965,
|
| 56 |
+
"eval_stsbenchmark_test_spearman": 0.5981378536949618,
|
| 57 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.6199557866619861,
|
| 58 |
+
"step": 1250
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"epoch": 0.12,
|
| 62 |
+
"learning_rate": 6.661926768574476e-06,
|
| 63 |
+
"loss": 0.0003,
|
| 64 |
+
"step": 1875
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"epoch": 0.12,
|
| 68 |
+
"eval_avg_sts": 0.6494555691959818,
|
| 69 |
+
"eval_avg_test": 0.6471419605134568,
|
| 70 |
+
"eval_avg_train_and_dev": 0.6641694696137254,
|
| 71 |
+
"eval_runtime": 21.9114,
|
| 72 |
+
"eval_sickrelatedness_spearman": 0.6689216651276976,
|
| 73 |
+
"eval_sickrelatedness_test_spearman": 0.6660745763942842,
|
| 74 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.6717733112444599,
|
| 75 |
+
"eval_sts12_spearman": 0.4998164334853574,
|
| 76 |
+
"eval_sts13_spearman": 0.7060812594730672,
|
| 77 |
+
"eval_sts14_spearman": 0.5974341417069697,
|
| 78 |
+
"eval_sts15_spearman": 0.7041580385169013,
|
| 79 |
+
"eval_sts16_spearman": 0.7158255254242747,
|
| 80 |
+
"eval_stsbenchmark_spearman": 0.6539519206376052,
|
| 81 |
+
"eval_stsbenchmark_test_spearman": 0.6406037485933435,
|
| 82 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.6565656279829908,
|
| 83 |
+
"step": 1875
|
| 84 |
+
},
|
| 85 |
+
{
|
| 86 |
+
"epoch": 0.16,
|
| 87 |
+
"learning_rate": 8.883753999289016e-06,
|
| 88 |
+
"loss": 0.0003,
|
| 89 |
+
"step": 2500
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"epoch": 0.16,
|
| 93 |
+
"eval_avg_sts": 0.6720172723077595,
|
| 94 |
+
"eval_avg_test": 0.6700901794645714,
|
| 95 |
+
"eval_avg_train_and_dev": 0.6813887988919534,
|
| 96 |
+
"eval_runtime": 21.6305,
|
| 97 |
+
"eval_sickrelatedness_spearman": 0.6752369370120973,
|
| 98 |
+
"eval_sickrelatedness_test_spearman": 0.6720545832463273,
|
| 99 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.6784163347898827,
|
| 100 |
+
"eval_sts12_spearman": 0.5357838277332643,
|
| 101 |
+
"eval_sts13_spearman": 0.724945557702781,
|
| 102 |
+
"eval_sts14_spearman": 0.6231065986339928,
|
| 103 |
+
"eval_sts15_spearman": 0.7325580598726732,
|
| 104 |
+
"eval_sts16_spearman": 0.7302626467999829,
|
| 105 |
+
"eval_stsbenchmark_spearman": 0.6822272783995248,
|
| 106 |
+
"eval_stsbenchmark_test_spearman": 0.6719199822629783,
|
| 107 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.6843612629940239,
|
| 108 |
+
"step": 2500
|
| 109 |
+
},
|
| 110 |
+
{
|
| 111 |
+
"epoch": 0.2,
|
| 112 |
+
"learning_rate": 9.929417638781718e-06,
|
| 113 |
+
"loss": 0.0002,
|
| 114 |
+
"step": 3125
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"epoch": 0.2,
|
| 118 |
+
"eval_avg_sts": 0.6909069613326249,
|
| 119 |
+
"eval_avg_test": 0.6892152885857711,
|
| 120 |
+
"eval_avg_train_and_dev": 0.6981092005408447,
|
| 121 |
+
"eval_runtime": 21.3949,
|
| 122 |
+
"eval_sickrelatedness_spearman": 0.684142432010355,
|
| 123 |
+
"eval_sickrelatedness_test_spearman": 0.6810570206913322,
|
| 124 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.6872006477750381,
|
| 125 |
+
"eval_sts12_spearman": 0.571210204194906,
|
| 126 |
+
"eval_sts13_spearman": 0.744024983013372,
|
| 127 |
+
"eval_sts14_spearman": 0.6427295960479041,
|
| 128 |
+
"eval_sts15_spearman": 0.75072362281639,
|
| 129 |
+
"eval_sts16_spearman": 0.7362828177459125,
|
| 130 |
+
"eval_stsbenchmark_spearman": 0.7072350734995339,
|
| 131 |
+
"eval_stsbenchmark_test_spearman": 0.6984787755905806,
|
| 132 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7090177533066512,
|
| 133 |
+
"step": 3125
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"epoch": 0.24,
|
| 137 |
+
"learning_rate": 9.787572057555263e-06,
|
| 138 |
+
"loss": 0.0001,
|
| 139 |
+
"step": 3750
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"epoch": 0.24,
|
| 143 |
+
"eval_avg_sts": 0.7110034740131804,
|
| 144 |
+
"eval_avg_test": 0.7090073777235183,
|
| 145 |
+
"eval_avg_train_and_dev": 0.7152562951128082,
|
| 146 |
+
"eval_runtime": 21.4902,
|
| 147 |
+
"eval_sickrelatedness_spearman": 0.6924481449933352,
|
| 148 |
+
"eval_sickrelatedness_test_spearman": 0.6896546207158979,
|
| 149 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.6953282368083006,
|
| 150 |
+
"eval_sts12_spearman": 0.5925559956254365,
|
| 151 |
+
"eval_sts13_spearman": 0.7670195193741965,
|
| 152 |
+
"eval_sts14_spearman": 0.6676224547832722,
|
| 153 |
+
"eval_sts15_spearman": 0.7714028016335752,
|
| 154 |
+
"eval_sts16_spearman": 0.7530593474380043,
|
| 155 |
+
"eval_stsbenchmark_spearman": 0.7329160542444434,
|
| 156 |
+
"eval_stsbenchmark_test_spearman": 0.7217369044942459,
|
| 157 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7351843534173158,
|
| 158 |
+
"step": 3750
|
| 159 |
+
},
|
| 160 |
+
{
|
| 161 |
+
"epoch": 0.28,
|
| 162 |
+
"learning_rate": 9.645726476328811e-06,
|
| 163 |
+
"loss": 0.0002,
|
| 164 |
+
"step": 4375
|
| 165 |
+
},
|
| 166 |
+
{
|
| 167 |
+
"epoch": 0.28,
|
| 168 |
+
"eval_avg_sts": 0.7194580287637145,
|
| 169 |
+
"eval_avg_test": 0.717033045368064,
|
| 170 |
+
"eval_avg_train_and_dev": 0.72525371399646,
|
| 171 |
+
"eval_runtime": 21.832,
|
| 172 |
+
"eval_sickrelatedness_spearman": 0.69964732104727,
|
| 173 |
+
"eval_sickrelatedness_test_spearman": 0.6974584549846506,
|
| 174 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7019540331749327,
|
| 175 |
+
"eval_sts12_spearman": 0.6087565607405551,
|
| 176 |
+
"eval_sts13_spearman": 0.775181610329269,
|
| 177 |
+
"eval_sts14_spearman": 0.6807344815935793,
|
| 178 |
+
"eval_sts15_spearman": 0.7714449656594323,
|
| 179 |
+
"eval_sts16_spearman": 0.7549091580553093,
|
| 180 |
+
"eval_stsbenchmark_spearman": 0.7455321039205858,
|
| 181 |
+
"eval_stsbenchmark_test_spearman": 0.7307460862136526,
|
| 182 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7485533948179873,
|
| 183 |
+
"step": 4375
|
| 184 |
+
},
|
| 185 |
+
{
|
| 186 |
+
"epoch": 0.32,
|
| 187 |
+
"learning_rate": 9.503880895102356e-06,
|
| 188 |
+
"loss": 0.0002,
|
| 189 |
+
"step": 5000
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"epoch": 0.32,
|
| 193 |
+
"eval_avg_sts": 0.7357926871862974,
|
| 194 |
+
"eval_avg_test": 0.7337295361557176,
|
| 195 |
+
"eval_avg_train_and_dev": 0.7387626368976384,
|
| 196 |
+
"eval_runtime": 21.3496,
|
| 197 |
+
"eval_sickrelatedness_spearman": 0.7081277931755735,
|
| 198 |
+
"eval_sickrelatedness_test_spearman": 0.7070731150129818,
|
| 199 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7093962696914289,
|
| 200 |
+
"eval_sts12_spearman": 0.6344626415135579,
|
| 201 |
+
"eval_sts13_spearman": 0.7877078434493416,
|
| 202 |
+
"eval_sts14_spearman": 0.6986755801629566,
|
| 203 |
+
"eval_sts15_spearman": 0.7910799783582211,
|
| 204 |
+
"eval_sts16_spearman": 0.7651327511753051,
|
| 205 |
+
"eval_stsbenchmark_spearman": 0.7653622224691259,
|
| 206 |
+
"eval_stsbenchmark_test_spearman": 0.7519748434176595,
|
| 207 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.768129004103848,
|
| 208 |
+
"step": 5000
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"epoch": 0.36,
|
| 212 |
+
"learning_rate": 9.362035313875902e-06,
|
| 213 |
+
"loss": 0.0001,
|
| 214 |
+
"step": 5625
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"epoch": 0.36,
|
| 218 |
+
"eval_avg_sts": 0.7418773860073935,
|
| 219 |
+
"eval_avg_test": 0.7395590037062557,
|
| 220 |
+
"eval_avg_train_and_dev": 0.743014478903171,
|
| 221 |
+
"eval_runtime": 22.6753,
|
| 222 |
+
"eval_sickrelatedness_spearman": 0.7091920474244533,
|
| 223 |
+
"eval_sickrelatedness_test_spearman": 0.7076497343346897,
|
| 224 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7108735600549281,
|
| 225 |
+
"eval_sts12_spearman": 0.6459740297105186,
|
| 226 |
+
"eval_sts13_spearman": 0.7946896206600376,
|
| 227 |
+
"eval_sts14_spearman": 0.706190273571107,
|
| 228 |
+
"eval_sts15_spearman": 0.7965399044791899,
|
| 229 |
+
"eval_sts16_spearman": 0.7684287092404692,
|
| 230 |
+
"eval_stsbenchmark_spearman": 0.7721271169659789,
|
| 231 |
+
"eval_stsbenchmark_test_spearman": 0.7574407539477783,
|
| 232 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7751553977514137,
|
| 233 |
+
"step": 5625
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"epoch": 0.4,
|
| 237 |
+
"learning_rate": 9.220189732649449e-06,
|
| 238 |
+
"loss": 0.0001,
|
| 239 |
+
"step": 6250
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"epoch": 0.4,
|
| 243 |
+
"eval_avg_sts": 0.7308385327663487,
|
| 244 |
+
"eval_avg_test": 0.7283863218628899,
|
| 245 |
+
"eval_avg_train_and_dev": 0.7345345112957317,
|
| 246 |
+
"eval_runtime": 21.8169,
|
| 247 |
+
"eval_sickrelatedness_spearman": 0.7046853908818935,
|
| 248 |
+
"eval_sickrelatedness_test_spearman": 0.7037265532974459,
|
| 249 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7057337293264881,
|
| 250 |
+
"eval_sts12_spearman": 0.621875598643423,
|
| 251 |
+
"eval_sts13_spearman": 0.7860303924520814,
|
| 252 |
+
"eval_sts14_spearman": 0.694793783643542,
|
| 253 |
+
"eval_sts15_spearman": 0.786890544888634,
|
| 254 |
+
"eval_sts16_spearman": 0.761570793834228,
|
| 255 |
+
"eval_stsbenchmark_spearman": 0.7600232250206383,
|
| 256 |
+
"eval_stsbenchmark_test_spearman": 0.7438165862808751,
|
| 257 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7633352932649752,
|
| 258 |
+
"step": 6250
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"epoch": 0.44,
|
| 262 |
+
"learning_rate": 9.078344151422995e-06,
|
| 263 |
+
"loss": 0.0002,
|
| 264 |
+
"step": 6875
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"epoch": 0.44,
|
| 268 |
+
"eval_avg_sts": 0.7309661390558988,
|
| 269 |
+
"eval_avg_test": 0.7286727854196379,
|
| 270 |
+
"eval_avg_train_and_dev": 0.7359208652832265,
|
| 271 |
+
"eval_runtime": 21.1351,
|
| 272 |
+
"eval_sickrelatedness_spearman": 0.7048537796254629,
|
| 273 |
+
"eval_sickrelatedness_test_spearman": 0.702412783558349,
|
| 274 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7073541509053075,
|
| 275 |
+
"eval_sts12_spearman": 0.6186911062602717,
|
| 276 |
+
"eval_sts13_spearman": 0.7856611011058849,
|
| 277 |
+
"eval_sts14_spearman": 0.6972845097860244,
|
| 278 |
+
"eval_sts15_spearman": 0.7842731788758606,
|
| 279 |
+
"eval_sts16_spearman": 0.7642482925335029,
|
| 280 |
+
"eval_stsbenchmark_spearman": 0.761751005204284,
|
| 281 |
+
"eval_stsbenchmark_test_spearman": 0.7481385258175712,
|
| 282 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7644875796611457,
|
| 283 |
+
"step": 6875
|
| 284 |
+
},
|
| 285 |
+
{
|
| 286 |
+
"epoch": 0.48,
|
| 287 |
+
"learning_rate": 8.936498570196542e-06,
|
| 288 |
+
"loss": 0.0001,
|
| 289 |
+
"step": 7500
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"epoch": 0.48,
|
| 293 |
+
"eval_avg_sts": 0.7398140587015002,
|
| 294 |
+
"eval_avg_test": 0.7375789283159258,
|
| 295 |
+
"eval_avg_train_and_dev": 0.7436463146851804,
|
| 296 |
+
"eval_runtime": 21.5092,
|
| 297 |
+
"eval_sickrelatedness_spearman": 0.7104350424990357,
|
| 298 |
+
"eval_sickrelatedness_test_spearman": 0.707645634808155,
|
| 299 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7133276839990611,
|
| 300 |
+
"eval_sts12_spearman": 0.6311869476757046,
|
| 301 |
+
"eval_sts13_spearman": 0.7943880897002324,
|
| 302 |
+
"eval_sts14_spearman": 0.7054140851208188,
|
| 303 |
+
"eval_sts15_spearman": 0.7965258267671331,
|
| 304 |
+
"eval_sts16_spearman": 0.7693214228482549,
|
| 305 |
+
"eval_stsbenchmark_spearman": 0.7714269962993212,
|
| 306 |
+
"eval_stsbenchmark_test_spearman": 0.7585704912911815,
|
| 307 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7739649453712996,
|
| 308 |
+
"step": 7500
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"epoch": 0.52,
|
| 312 |
+
"learning_rate": 8.794652988970088e-06,
|
| 313 |
+
"loss": 0.0001,
|
| 314 |
+
"step": 8125
|
| 315 |
+
},
|
| 316 |
+
{
|
| 317 |
+
"epoch": 0.52,
|
| 318 |
+
"eval_avg_sts": 0.7373193093280432,
|
| 319 |
+
"eval_avg_test": 0.7349110835074212,
|
| 320 |
+
"eval_avg_train_and_dev": 0.74065035952663,
|
| 321 |
+
"eval_runtime": 21.2362,
|
| 322 |
+
"eval_sickrelatedness_spearman": 0.7089799957481375,
|
| 323 |
+
"eval_sickrelatedness_test_spearman": 0.7065719679366725,
|
| 324 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.711364735604823,
|
| 325 |
+
"eval_sts12_spearman": 0.62980351920814,
|
| 326 |
+
"eval_sts13_spearman": 0.7927297021563888,
|
| 327 |
+
"eval_sts14_spearman": 0.7002451810649419,
|
| 328 |
+
"eval_sts15_spearman": 0.7943326132694016,
|
| 329 |
+
"eval_sts16_spearman": 0.7680353970909194,
|
| 330 |
+
"eval_stsbenchmark_spearman": 0.767108756758373,
|
| 331 |
+
"eval_stsbenchmark_test_spearman": 0.7526592038254838,
|
| 332 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.769935983448437,
|
| 333 |
+
"step": 8125
|
| 334 |
+
},
|
| 335 |
+
{
|
| 336 |
+
"epoch": 0.56,
|
| 337 |
+
"learning_rate": 8.652807407743635e-06,
|
| 338 |
+
"loss": 0.0001,
|
| 339 |
+
"step": 8750
|
| 340 |
+
},
|
| 341 |
+
{
|
| 342 |
+
"epoch": 0.56,
|
| 343 |
+
"eval_avg_sts": 0.7363613158654552,
|
| 344 |
+
"eval_avg_test": 0.7344457311029845,
|
| 345 |
+
"eval_avg_train_and_dev": 0.7385101274896235,
|
| 346 |
+
"eval_runtime": 21.007,
|
| 347 |
+
"eval_sickrelatedness_spearman": 0.7079709220981268,
|
| 348 |
+
"eval_sickrelatedness_test_spearman": 0.7060740065988659,
|
| 349 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7098373581622875,
|
| 350 |
+
"eval_sts12_spearman": 0.6245604819610331,
|
| 351 |
+
"eval_sts13_spearman": 0.7940661258079552,
|
| 352 |
+
"eval_sts14_spearman": 0.6995530567749194,
|
| 353 |
+
"eval_sts15_spearman": 0.7934026178932524,
|
| 354 |
+
"eval_sts16_spearman": 0.7701123683553529,
|
| 355 |
+
"eval_stsbenchmark_spearman": 0.764863638167547,
|
| 356 |
+
"eval_stsbenchmark_test_spearman": 0.7533514603295133,
|
| 357 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7671828968169594,
|
| 358 |
+
"step": 8750
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"epoch": 0.6,
|
| 362 |
+
"learning_rate": 8.510961826517181e-06,
|
| 363 |
+
"loss": 0.0001,
|
| 364 |
+
"step": 9375
|
| 365 |
+
},
|
| 366 |
+
{
|
| 367 |
+
"epoch": 0.6,
|
| 368 |
+
"eval_avg_sts": 0.7318749663720026,
|
| 369 |
+
"eval_avg_test": 0.7298439077988678,
|
| 370 |
+
"eval_avg_train_and_dev": 0.7324391542106021,
|
| 371 |
+
"eval_runtime": 21.8218,
|
| 372 |
+
"eval_sickrelatedness_spearman": 0.7058228056610777,
|
| 373 |
+
"eval_sickrelatedness_test_spearman": 0.7039121780620011,
|
| 374 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7077349771782333,
|
| 375 |
+
"eval_sts12_spearman": 0.618700449208015,
|
| 376 |
+
"eval_sts13_spearman": 0.7954576133369399,
|
| 377 |
+
"eval_sts14_spearman": 0.6962032852506622,
|
| 378 |
+
"eval_sts15_spearman": 0.7872116068915151,
|
| 379 |
+
"eval_sts16_spearman": 0.765104808041975,
|
| 380 |
+
"eval_stsbenchmark_spearman": 0.7546241962138335,
|
| 381 |
+
"eval_stsbenchmark_test_spearman": 0.7423174138009657,
|
| 382 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7571433312429708,
|
| 383 |
+
"step": 9375
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"epoch": 0.64,
|
| 387 |
+
"learning_rate": 8.369116245290728e-06,
|
| 388 |
+
"loss": 0.0001,
|
| 389 |
+
"step": 10000
|
| 390 |
+
},
|
| 391 |
+
{
|
| 392 |
+
"epoch": 0.64,
|
| 393 |
+
"eval_avg_sts": 0.7409280134068449,
|
| 394 |
+
"eval_avg_test": 0.7391325218596417,
|
| 395 |
+
"eval_avg_train_and_dev": 0.7415103794141908,
|
| 396 |
+
"eval_runtime": 21.4143,
|
| 397 |
+
"eval_sickrelatedness_spearman": 0.7104368210609439,
|
| 398 |
+
"eval_sickrelatedness_test_spearman": 0.7079528143535209,
|
| 399 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7129369206822436,
|
| 400 |
+
"eval_sts12_spearman": 0.6369900901886509,
|
| 401 |
+
"eval_sts13_spearman": 0.8003649874497492,
|
| 402 |
+
"eval_sts14_spearman": 0.7048130149249648,
|
| 403 |
+
"eval_sts15_spearman": 0.7956828867682607,
|
| 404 |
+
"eval_sts16_spearman": 0.7700745630771816,
|
| 405 |
+
"eval_stsbenchmark_spearman": 0.7681337303781637,
|
| 406 |
+
"eval_stsbenchmark_test_spearman": 0.7580492962551647,
|
| 407 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.770083838146138,
|
| 408 |
+
"step": 10000
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"epoch": 0.68,
|
| 412 |
+
"learning_rate": 8.227497616994236e-06,
|
| 413 |
+
"loss": 0.0003,
|
| 414 |
+
"step": 10625
|
| 415 |
+
},
|
| 416 |
+
{
|
| 417 |
+
"epoch": 0.68,
|
| 418 |
+
"eval_avg_sts": 0.7359543470984595,
|
| 419 |
+
"eval_avg_test": 0.733931880207311,
|
| 420 |
+
"eval_avg_train_and_dev": 0.7367064905162184,
|
| 421 |
+
"eval_runtime": 21.2567,
|
| 422 |
+
"eval_sickrelatedness_spearman": 0.7081098389229429,
|
| 423 |
+
"eval_sickrelatedness_test_spearman": 0.7069070856977242,
|
| 424 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7093919257525952,
|
| 425 |
+
"eval_sts12_spearman": 0.6171012791250989,
|
| 426 |
+
"eval_sts13_spearman": 0.8001733717764794,
|
| 427 |
+
"eval_sts14_spearman": 0.7010514936978448,
|
| 428 |
+
"eval_sts15_spearman": 0.7914208067428021,
|
| 429 |
+
"eval_sts16_spearman": 0.7723885383628244,
|
| 430 |
+
"eval_stsbenchmark_spearman": 0.7614351010612246,
|
| 431 |
+
"eval_stsbenchmark_test_spearman": 0.7484805860484034,
|
| 432 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7640210552798415,
|
| 433 |
+
"step": 10625
|
| 434 |
+
},
|
| 435 |
+
{
|
| 436 |
+
"epoch": 0.72,
|
| 437 |
+
"learning_rate": 8.085652035767782e-06,
|
| 438 |
+
"loss": 0.0002,
|
| 439 |
+
"step": 11250
|
| 440 |
+
},
|
| 441 |
+
{
|
| 442 |
+
"epoch": 0.72,
|
| 443 |
+
"eval_avg_sts": 0.7331832920502,
|
| 444 |
+
"eval_avg_test": 0.7296440457200707,
|
| 445 |
+
"eval_avg_train_and_dev": 0.7365755808733788,
|
| 446 |
+
"eval_runtime": 22.0032,
|
| 447 |
+
"eval_sickrelatedness_spearman": 0.7101591391042099,
|
| 448 |
+
"eval_sickrelatedness_test_spearman": 0.7080075712587889,
|
| 449 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7123952349576758,
|
| 450 |
+
"eval_sts12_spearman": 0.6071610926796485,
|
| 451 |
+
"eval_sts13_spearman": 0.7986426077282267,
|
| 452 |
+
"eval_sts14_spearman": 0.7048385507770369,
|
| 453 |
+
"eval_sts15_spearman": 0.790486863635161,
|
| 454 |
+
"eval_sts16_spearman": 0.7648214181893283,
|
| 455 |
+
"eval_stsbenchmark_spearman": 0.7561733722377888,
|
| 456 |
+
"eval_stsbenchmark_test_spearman": 0.7335502157723038,
|
| 457 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7607559267890819,
|
| 458 |
+
"step": 11250
|
| 459 |
+
},
|
| 460 |
+
{
|
| 461 |
+
"epoch": 0.76,
|
| 462 |
+
"learning_rate": 7.943806454541329e-06,
|
| 463 |
+
"loss": 0.0001,
|
| 464 |
+
"step": 11875
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"epoch": 0.76,
|
| 468 |
+
"eval_avg_sts": 0.7284741669636894,
|
| 469 |
+
"eval_avg_test": 0.7249847364931884,
|
| 470 |
+
"eval_avg_train_and_dev": 0.7313381333368812,
|
| 471 |
+
"eval_runtime": 21.7362,
|
| 472 |
+
"eval_sickrelatedness_spearman": 0.7051030572507782,
|
| 473 |
+
"eval_sickrelatedness_test_spearman": 0.7020350055407318,
|
| 474 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7082106452571797,
|
| 475 |
+
"eval_sts12_spearman": 0.5997827794595306,
|
| 476 |
+
"eval_sts13_spearman": 0.7980496603077974,
|
| 477 |
+
"eval_sts14_spearman": 0.6948829895910905,
|
| 478 |
+
"eval_sts15_spearman": 0.7853112805674828,
|
| 479 |
+
"eval_sts16_spearman": 0.7659372500003789,
|
| 480 |
+
"eval_stsbenchmark_spearman": 0.7502521515687675,
|
| 481 |
+
"eval_stsbenchmark_test_spearman": 0.728894189985307,
|
| 482 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7544656214165828,
|
| 483 |
+
"step": 11875
|
| 484 |
+
},
|
| 485 |
+
{
|
| 486 |
+
"epoch": 0.8,
|
| 487 |
+
"learning_rate": 7.802187826244837e-06,
|
| 488 |
+
"loss": 0.0002,
|
| 489 |
+
"step": 12500
|
| 490 |
+
},
|
| 491 |
+
{
|
| 492 |
+
"epoch": 0.8,
|
| 493 |
+
"eval_avg_sts": 0.7330359711356126,
|
| 494 |
+
"eval_avg_test": 0.7291193278783638,
|
| 495 |
+
"eval_avg_train_and_dev": 0.735211801074501,
|
| 496 |
+
"eval_runtime": 20.8416,
|
| 497 |
+
"eval_sickrelatedness_spearman": 0.7097679817259296,
|
| 498 |
+
"eval_sickrelatedness_test_spearman": 0.7055788799611423,
|
| 499 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7139001197937711,
|
| 500 |
+
"eval_sts12_spearman": 0.6106517927738067,
|
| 501 |
+
"eval_sts13_spearman": 0.801722782519613,
|
| 502 |
+
"eval_sts14_spearman": 0.697830283597753,
|
| 503 |
+
"eval_sts15_spearman": 0.7905626238832419,
|
| 504 |
+
"eval_sts16_spearman": 0.7686769836942824,
|
| 505 |
+
"eval_stsbenchmark_spearman": 0.7520393497546612,
|
| 506 |
+
"eval_stsbenchmark_test_spearman": 0.7288119487187074,
|
| 507 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7565234823552308,
|
| 508 |
+
"step": 12500
|
| 509 |
+
},
|
| 510 |
+
{
|
| 511 |
+
"epoch": 0.84,
|
| 512 |
+
"learning_rate": 7.660342245018383e-06,
|
| 513 |
+
"loss": 0.0001,
|
| 514 |
+
"step": 13125
|
| 515 |
+
},
|
| 516 |
+
{
|
| 517 |
+
"epoch": 0.84,
|
| 518 |
+
"eval_avg_sts": 0.7378956830798117,
|
| 519 |
+
"eval_avg_test": 0.733899347554639,
|
| 520 |
+
"eval_avg_train_and_dev": 0.7387413140477561,
|
| 521 |
+
"eval_runtime": 21.015,
|
| 522 |
+
"eval_sickrelatedness_spearman": 0.710929018302664,
|
| 523 |
+
"eval_sickrelatedness_test_spearman": 0.7063428540695924,
|
| 524 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7154190994394085,
|
| 525 |
+
"eval_sts12_spearman": 0.6233547440533763,
|
| 526 |
+
"eval_sts13_spearman": 0.8057484430966548,
|
| 527 |
+
"eval_sts14_spearman": 0.7064661836743785,
|
| 528 |
+
"eval_sts15_spearman": 0.788799289281052,
|
| 529 |
+
"eval_sts16_spearman": 0.7725484698411716,
|
| 530 |
+
"eval_stsbenchmark_spearman": 0.7574236333093849,
|
| 531 |
+
"eval_stsbenchmark_test_spearman": 0.7340354488662467,
|
| 532 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7620635286561037,
|
| 533 |
+
"step": 13125
|
| 534 |
+
},
|
| 535 |
+
{
|
| 536 |
+
"epoch": 0.88,
|
| 537 |
+
"learning_rate": 7.51849666379193e-06,
|
| 538 |
+
"loss": 0.0001,
|
| 539 |
+
"step": 13750
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"epoch": 0.88,
|
| 543 |
+
"eval_avg_sts": 0.7378627077777866,
|
| 544 |
+
"eval_avg_test": 0.7339054575865804,
|
| 545 |
+
"eval_avg_train_and_dev": 0.7381744196853126,
|
| 546 |
+
"eval_runtime": 20.944,
|
| 547 |
+
"eval_sickrelatedness_spearman": 0.7089921810762922,
|
| 548 |
+
"eval_sickrelatedness_test_spearman": 0.70509102387014,
|
| 549 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7128834324759727,
|
| 550 |
+
"eval_sts12_spearman": 0.6228312321043947,
|
| 551 |
+
"eval_sts13_spearman": 0.8058975575766927,
|
| 552 |
+
"eval_sts14_spearman": 0.7070969399502143,
|
| 553 |
+
"eval_sts15_spearman": 0.7893837860149883,
|
| 554 |
+
"eval_sts16_spearman": 0.7720301633604001,
|
| 555 |
+
"eval_stsbenchmark_spearman": 0.7588070943615248,
|
| 556 |
+
"eval_stsbenchmark_test_spearman": 0.7350075002292331,
|
| 557 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7634654068946525,
|
| 558 |
+
"step": 13750
|
| 559 |
+
},
|
| 560 |
+
{
|
| 561 |
+
"epoch": 0.92,
|
| 562 |
+
"learning_rate": 7.376651082565477e-06,
|
| 563 |
+
"loss": 0.0001,
|
| 564 |
+
"step": 14375
|
| 565 |
+
},
|
| 566 |
+
{
|
| 567 |
+
"epoch": 0.92,
|
| 568 |
+
"eval_avg_sts": 0.7359929925111067,
|
| 569 |
+
"eval_avg_test": 0.7323951374834617,
|
| 570 |
+
"eval_avg_train_and_dev": 0.7367857943216309,
|
| 571 |
+
"eval_runtime": 20.9186,
|
| 572 |
+
"eval_sickrelatedness_spearman": 0.7088303588545686,
|
| 573 |
+
"eval_sickrelatedness_test_spearman": 0.7063407570964759,
|
| 574 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7113763758675955,
|
| 575 |
+
"eval_sts12_spearman": 0.6136120965617352,
|
| 576 |
+
"eval_sts13_spearman": 0.8054760021735932,
|
| 577 |
+
"eval_sts14_spearman": 0.7046761197182577,
|
| 578 |
+
"eval_sts15_spearman": 0.7905370837297951,
|
| 579 |
+
"eval_sts16_spearman": 0.7711103256070158,
|
| 580 |
+
"eval_stsbenchmark_spearman": 0.7577089609327817,
|
| 581 |
+
"eval_stsbenchmark_test_spearman": 0.7350135774973598,
|
| 582 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7621952127756663,
|
| 583 |
+
"step": 14375
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"epoch": 0.96,
|
| 587 |
+
"learning_rate": 7.2348055013390225e-06,
|
| 588 |
+
"loss": 0.0001,
|
| 589 |
+
"step": 15000
|
| 590 |
+
},
|
| 591 |
+
{
|
| 592 |
+
"epoch": 0.96,
|
| 593 |
+
"eval_avg_sts": 0.7361943130200425,
|
| 594 |
+
"eval_avg_test": 0.7321974439132747,
|
| 595 |
+
"eval_avg_train_and_dev": 0.7382189598774902,
|
| 596 |
+
"eval_runtime": 21.1353,
|
| 597 |
+
"eval_sickrelatedness_spearman": 0.7096126852616816,
|
| 598 |
+
"eval_sickrelatedness_test_spearman": 0.7056509931468986,
|
| 599 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7134760429404705,
|
| 600 |
+
"eval_sts12_spearman": 0.613133445817732,
|
| 601 |
+
"eval_sts13_spearman": 0.8048892508629485,
|
| 602 |
+
"eval_sts14_spearman": 0.7046336569955289,
|
| 603 |
+
"eval_sts15_spearman": 0.7904009982047565,
|
| 604 |
+
"eval_sts16_spearman": 0.7725216175809143,
|
| 605 |
+
"eval_stsbenchmark_spearman": 0.7581685364167358,
|
| 606 |
+
"eval_stsbenchmark_test_spearman": 0.7341521447841448,
|
| 607 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7629618768145098,
|
| 608 |
+
"step": 15000
|
| 609 |
+
},
|
| 610 |
+
{
|
| 611 |
+
"epoch": 1.0,
|
| 612 |
+
"learning_rate": 7.092959920112569e-06,
|
| 613 |
+
"loss": 0.0001,
|
| 614 |
+
"step": 15625
|
| 615 |
+
},
|
| 616 |
+
{
|
| 617 |
+
"epoch": 1.0,
|
| 618 |
+
"eval_avg_sts": 0.7375598442201905,
|
| 619 |
+
"eval_avg_test": 0.7335423238398006,
|
| 620 |
+
"eval_avg_train_and_dev": 0.7395439996671019,
|
| 621 |
+
"eval_runtime": 20.8993,
|
| 622 |
+
"eval_sickrelatedness_spearman": 0.707283647594327,
|
| 623 |
+
"eval_sickrelatedness_test_spearman": 0.703616122718381,
|
| 624 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7109751567220617,
|
| 625 |
+
"eval_sts12_spearman": 0.6159411441163902,
|
| 626 |
+
"eval_sts13_spearman": 0.8049060590081953,
|
| 627 |
+
"eval_sts14_spearman": 0.7108002254446067,
|
| 628 |
+
"eval_sts15_spearman": 0.788937840589697,
|
| 629 |
+
"eval_sts16_spearman": 0.7718528831999566,
|
| 630 |
+
"eval_stsbenchmark_spearman": 0.7631971095881599,
|
| 631 |
+
"eval_stsbenchmark_test_spearman": 0.7387419918013778,
|
| 632 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7681128426121422,
|
| 633 |
+
"step": 15625
|
| 634 |
+
},
|
| 635 |
+
{
|
| 636 |
+
"epoch": 1.04,
|
| 637 |
+
"learning_rate": 6.951114338886115e-06,
|
| 638 |
+
"loss": 0.0001,
|
| 639 |
+
"step": 16250
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"epoch": 1.04,
|
| 643 |
+
"eval_avg_sts": 0.739199635865537,
|
| 644 |
+
"eval_avg_test": 0.7355897759739543,
|
| 645 |
+
"eval_avg_train_and_dev": 0.739808186948654,
|
| 646 |
+
"eval_runtime": 20.9838,
|
| 647 |
+
"eval_sickrelatedness_spearman": 0.7080971683598302,
|
| 648 |
+
"eval_sickrelatedness_test_spearman": 0.7051237540095253,
|
| 649 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7111032002738162,
|
| 650 |
+
"eval_sts12_spearman": 0.6194643102288157,
|
| 651 |
+
"eval_sts13_spearman": 0.8098252612446439,
|
| 652 |
+
"eval_sts14_spearman": 0.7116464283311791,
|
| 653 |
+
"eval_sts15_spearman": 0.7899553665444689,
|
| 654 |
+
"eval_sts16_spearman": 0.7712889077310251,
|
| 655 |
+
"eval_stsbenchmark_spearman": 0.7641200086187968,
|
| 656 |
+
"eval_stsbenchmark_test_spearman": 0.7418244037280217,
|
| 657 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7685131736234918,
|
| 658 |
+
"step": 16250
|
| 659 |
+
},
|
| 660 |
+
{
|
| 661 |
+
"epoch": 1.08,
|
| 662 |
+
"learning_rate": 6.809495710589624e-06,
|
| 663 |
+
"loss": 0.0001,
|
| 664 |
+
"step": 16875
|
| 665 |
+
},
|
| 666 |
+
{
|
| 667 |
+
"epoch": 1.08,
|
| 668 |
+
"eval_avg_sts": 0.7466792846411037,
|
| 669 |
+
"eval_avg_test": 0.7432402664611477,
|
| 670 |
+
"eval_avg_train_and_dev": 0.7444846866291048,
|
| 671 |
+
"eval_runtime": 21.2164,
|
| 672 |
+
"eval_sickrelatedness_spearman": 0.7071507310982136,
|
| 673 |
+
"eval_sickrelatedness_test_spearman": 0.7036683201355792,
|
| 674 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7106803741065932,
|
| 675 |
+
"eval_sts12_spearman": 0.6347617000535006,
|
| 676 |
+
"eval_sts13_spearman": 0.8154858577332686,
|
| 677 |
+
"eval_sts14_spearman": 0.7214692213354003,
|
| 678 |
+
"eval_sts15_spearman": 0.7972123472428747,
|
| 679 |
+
"eval_sts16_spearman": 0.7764992733505218,
|
| 680 |
+
"eval_stsbenchmark_spearman": 0.7741758616739455,
|
| 681 |
+
"eval_stsbenchmark_test_spearman": 0.7535851453768893,
|
| 682 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7782889991516164,
|
| 683 |
+
"step": 16875
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"epoch": 1.12,
|
| 687 |
+
"learning_rate": 6.66765012936317e-06,
|
| 688 |
+
"loss": 0.0,
|
| 689 |
+
"step": 17500
|
| 690 |
+
},
|
| 691 |
+
{
|
| 692 |
+
"epoch": 1.12,
|
| 693 |
+
"eval_avg_sts": 0.7445033109622727,
|
| 694 |
+
"eval_avg_test": 0.7410841678179132,
|
| 695 |
+
"eval_avg_train_and_dev": 0.7439110348985495,
|
| 696 |
+
"eval_runtime": 20.7911,
|
| 697 |
+
"eval_sickrelatedness_spearman": 0.7091854822762941,
|
| 698 |
+
"eval_sickrelatedness_test_spearman": 0.7062734147342097,
|
| 699 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7121887413339406,
|
| 700 |
+
"eval_sts12_spearman": 0.6279026674987728,
|
| 701 |
+
"eval_sts13_spearman": 0.8113067283249641,
|
| 702 |
+
"eval_sts14_spearman": 0.7213560371915237,
|
| 703 |
+
"eval_sts15_spearman": 0.7972695297709387,
|
| 704 |
+
"eval_sts16_spearman": 0.7730878646190888,
|
| 705 |
+
"eval_stsbenchmark_spearman": 0.7714148670543273,
|
| 706 |
+
"eval_stsbenchmark_test_spearman": 0.7503929325858947,
|
| 707 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7756333284631584,
|
| 708 |
+
"step": 17500
|
| 709 |
+
},
|
| 710 |
+
{
|
| 711 |
+
"epoch": 1.16,
|
| 712 |
+
"learning_rate": 6.52603150106668e-06,
|
| 713 |
+
"loss": 0.0001,
|
| 714 |
+
"step": 18125
|
| 715 |
+
},
|
| 716 |
+
{
|
| 717 |
+
"epoch": 1.16,
|
| 718 |
+
"eval_avg_sts": 0.7424258185781643,
|
| 719 |
+
"eval_avg_test": 0.7386605800972373,
|
| 720 |
+
"eval_avg_train_and_dev": 0.7454667750890833,
|
| 721 |
+
"eval_runtime": 21.0259,
|
| 722 |
+
"eval_sickrelatedness_spearman": 0.7126108043398651,
|
| 723 |
+
"eval_sickrelatedness_test_spearman": 0.7092073398325749,
|
| 724 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7160534091377377,
|
| 725 |
+
"eval_sts12_spearman": 0.6254715777237362,
|
| 726 |
+
"eval_sts13_spearman": 0.8088662475950285,
|
| 727 |
+
"eval_sts14_spearman": 0.7138096529469777,
|
| 728 |
+
"eval_sts15_spearman": 0.7925312951699506,
|
| 729 |
+
"eval_sts16_spearman": 0.7732924074022455,
|
| 730 |
+
"eval_stsbenchmark_spearman": 0.770398744869346,
|
| 731 |
+
"eval_stsbenchmark_test_spearman": 0.7474455400101482,
|
| 732 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7748801410404289,
|
| 733 |
+
"step": 18125
|
| 734 |
+
},
|
| 735 |
+
{
|
| 736 |
+
"epoch": 1.2,
|
| 737 |
+
"learning_rate": 6.384185919840225e-06,
|
| 738 |
+
"loss": 0.0001,
|
| 739 |
+
"step": 18750
|
| 740 |
+
},
|
| 741 |
+
{
|
| 742 |
+
"epoch": 1.2,
|
| 743 |
+
"eval_avg_sts": 0.7421874872103393,
|
| 744 |
+
"eval_avg_test": 0.7386332506855652,
|
| 745 |
+
"eval_avg_train_and_dev": 0.7448711406496367,
|
| 746 |
+
"eval_runtime": 21.0153,
|
| 747 |
+
"eval_sickrelatedness_spearman": 0.711420403952433,
|
| 748 |
+
"eval_sickrelatedness_test_spearman": 0.7078637859078531,
|
| 749 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7149714012084747,
|
| 750 |
+
"eval_sts12_spearman": 0.6198152391183793,
|
| 751 |
+
"eval_sts13_spearman": 0.8112609489937583,
|
| 752 |
+
"eval_sts14_spearman": 0.7137096793492043,
|
| 753 |
+
"eval_sts15_spearman": 0.7909819007518449,
|
| 754 |
+
"eval_sts16_spearman": 0.7775422386776029,
|
| 755 |
+
"eval_stsbenchmark_spearman": 0.7705819996291529,
|
| 756 |
+
"eval_stsbenchmark_test_spearman": 0.7492589620003135,
|
| 757 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7747708800907985,
|
| 758 |
+
"step": 18750
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"epoch": 1.24,
|
| 762 |
+
"learning_rate": 6.242340338613772e-06,
|
| 763 |
+
"loss": 0.0001,
|
| 764 |
+
"step": 19375
|
| 765 |
+
},
|
| 766 |
+
{
|
| 767 |
+
"epoch": 1.24,
|
| 768 |
+
"eval_avg_sts": 0.7433069618878057,
|
| 769 |
+
"eval_avg_test": 0.7395454400266834,
|
| 770 |
+
"eval_avg_train_and_dev": 0.7438536914759893,
|
| 771 |
+
"eval_runtime": 20.8728,
|
| 772 |
+
"eval_sickrelatedness_spearman": 0.7081170951655467,
|
| 773 |
+
"eval_sickrelatedness_test_spearman": 0.7032372294807069,
|
| 774 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7129498291505492,
|
| 775 |
+
"eval_sts12_spearman": 0.6319090861576243,
|
| 776 |
+
"eval_sts13_spearman": 0.8105293738494971,
|
| 777 |
+
"eval_sts14_spearman": 0.7133101665632953,
|
| 778 |
+
"eval_sts15_spearman": 0.7957949511234824,
|
| 779 |
+
"eval_sts16_spearman": 0.7729378020760672,
|
| 780 |
+
"eval_stsbenchmark_spearman": 0.7705502582791273,
|
| 781 |
+
"eval_stsbenchmark_test_spearman": 0.7490994709361111,
|
| 782 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7747575538014295,
|
| 783 |
+
"step": 19375
|
| 784 |
+
},
|
| 785 |
+
{
|
| 786 |
+
"epoch": 1.28,
|
| 787 |
+
"learning_rate": 6.100494757387318e-06,
|
| 788 |
+
"loss": 0.0001,
|
| 789 |
+
"step": 20000
|
| 790 |
+
},
|
| 791 |
+
{
|
| 792 |
+
"epoch": 1.28,
|
| 793 |
+
"eval_avg_sts": 0.7499060859363368,
|
| 794 |
+
"eval_avg_test": 0.7465428552421288,
|
| 795 |
+
"eval_avg_train_and_dev": 0.7475216864388343,
|
| 796 |
+
"eval_runtime": 20.8548,
|
| 797 |
+
"eval_sickrelatedness_spearman": 0.7077244255701862,
|
| 798 |
+
"eval_sickrelatedness_test_spearman": 0.7026718095093082,
|
| 799 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7128746659624327,
|
| 800 |
+
"eval_sts12_spearman": 0.6428011959930444,
|
| 801 |
+
"eval_sts13_spearman": 0.8168437631289807,
|
| 802 |
+
"eval_sts14_spearman": 0.7226103243605893,
|
| 803 |
+
"eval_sts15_spearman": 0.8024592797345551,
|
| 804 |
+
"eval_sts16_spearman": 0.778377683273261,
|
| 805 |
+
"eval_stsbenchmark_spearman": 0.778525929493741,
|
| 806 |
+
"eval_stsbenchmark_test_spearman": 0.7600359306951637,
|
| 807 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7821687069152359,
|
| 808 |
+
"step": 20000
|
| 809 |
+
},
|
| 810 |
+
{
|
| 811 |
+
"epoch": 1.32,
|
| 812 |
+
"learning_rate": 5.958649176160865e-06,
|
| 813 |
+
"loss": 0.0,
|
| 814 |
+
"step": 20625
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"epoch": 1.32,
|
| 818 |
+
"eval_avg_sts": 0.7484118293338831,
|
| 819 |
+
"eval_avg_test": 0.7449454204384204,
|
| 820 |
+
"eval_avg_train_and_dev": 0.7464359816566652,
|
| 821 |
+
"eval_runtime": 21.0711,
|
| 822 |
+
"eval_sickrelatedness_spearman": 0.7074503178963312,
|
| 823 |
+
"eval_sickrelatedness_test_spearman": 0.7023303881299484,
|
| 824 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7126512609959751,
|
| 825 |
+
"eval_sts12_spearman": 0.6387436970256535,
|
| 826 |
+
"eval_sts13_spearman": 0.8166418859715421,
|
| 827 |
+
"eval_sts14_spearman": 0.7212103904818994,
|
| 828 |
+
"eval_sts15_spearman": 0.8006237515601927,
|
| 829 |
+
"eval_sts16_spearman": 0.7777887546493993,
|
| 830 |
+
"eval_stsbenchmark_spearman": 0.7764240077521634,
|
| 831 |
+
"eval_stsbenchmark_test_spearman": 0.7572790752503075,
|
| 832 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7802207023173555,
|
| 833 |
+
"step": 20625
|
| 834 |
+
},
|
| 835 |
+
{
|
| 836 |
+
"epoch": 1.36,
|
| 837 |
+
"learning_rate": 5.816803594934411e-06,
|
| 838 |
+
"loss": 0.0002,
|
| 839 |
+
"step": 21250
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"epoch": 1.36,
|
| 843 |
+
"eval_avg_sts": 0.7350019199078851,
|
| 844 |
+
"eval_avg_test": 0.7318554174797746,
|
| 845 |
+
"eval_avg_train_and_dev": 0.7352898672690771,
|
| 846 |
+
"eval_runtime": 21.0366,
|
| 847 |
+
"eval_sickrelatedness_spearman": 0.7074999541454676,
|
| 848 |
+
"eval_sickrelatedness_test_spearman": 0.7046918538078386,
|
| 849 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7103029251438325,
|
| 850 |
+
"eval_sts12_spearman": 0.6156570708780759,
|
| 851 |
+
"eval_sts13_spearman": 0.8090436604472787,
|
| 852 |
+
"eval_sts14_spearman": 0.7028681750157657,
|
| 853 |
+
"eval_sts15_spearman": 0.7852035835160743,
|
| 854 |
+
"eval_sts16_spearman": 0.7682223919418893,
|
| 855 |
+
"eval_stsbenchmark_spearman": 0.7565186034106438,
|
| 856 |
+
"eval_stsbenchmark_test_spearman": 0.737301186751499,
|
| 857 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7602768093943217,
|
| 858 |
+
"step": 21250
|
| 859 |
+
},
|
| 860 |
+
{
|
| 861 |
+
"epoch": 1.4,
|
| 862 |
+
"learning_rate": 5.674958013707958e-06,
|
| 863 |
+
"loss": 0.0001,
|
| 864 |
+
"step": 21875
|
| 865 |
+
},
|
| 866 |
+
{
|
| 867 |
+
"epoch": 1.4,
|
| 868 |
+
"eval_avg_sts": 0.7380801801237084,
|
| 869 |
+
"eval_avg_test": 0.7352880148005531,
|
| 870 |
+
"eval_avg_train_and_dev": 0.7368474462537626,
|
| 871 |
+
"eval_runtime": 20.8923,
|
| 872 |
+
"eval_sickrelatedness_spearman": 0.7069077829822942,
|
| 873 |
+
"eval_sickrelatedness_test_spearman": 0.7031570262809338,
|
| 874 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7106689319787851,
|
| 875 |
+
"eval_sts12_spearman": 0.6231915518691485,
|
| 876 |
+
"eval_sts13_spearman": 0.812850852612454,
|
| 877 |
+
"eval_sts14_spearman": 0.7051573740637898,
|
| 878 |
+
"eval_sts15_spearman": 0.7880917747314226,
|
| 879 |
+
"eval_sts16_spearman": 0.7705225580148864,
|
| 880 |
+
"eval_stsbenchmark_spearman": 0.7598393665919628,
|
| 881 |
+
"eval_stsbenchmark_test_spearman": 0.7440449660312358,
|
| 882 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7630259605287402,
|
| 883 |
+
"step": 21875
|
| 884 |
+
},
|
| 885 |
+
{
|
| 886 |
+
"epoch": 1.44,
|
| 887 |
+
"learning_rate": 5.533112432481503e-06,
|
| 888 |
+
"loss": 0.0001,
|
| 889 |
+
"step": 22500
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"epoch": 1.44,
|
| 893 |
+
"eval_avg_sts": 0.7393794402273529,
|
| 894 |
+
"eval_avg_test": 0.7363579691480349,
|
| 895 |
+
"eval_avg_train_and_dev": 0.7395749061213661,
|
| 896 |
+
"eval_runtime": 21.2448,
|
| 897 |
+
"eval_sickrelatedness_spearman": 0.7102477930138079,
|
| 898 |
+
"eval_sickrelatedness_test_spearman": 0.7064793937973988,
|
| 899 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7139973441863221,
|
| 900 |
+
"eval_sts12_spearman": 0.6285417992352085,
|
| 901 |
+
"eval_sts13_spearman": 0.8114621677478028,
|
| 902 |
+
"eval_sts14_spearman": 0.7062564240963956,
|
| 903 |
+
"eval_sts15_spearman": 0.7886571693369652,
|
| 904 |
+
"eval_sts16_spearman": 0.7687956439175224,
|
| 905 |
+
"eval_stsbenchmark_spearman": 0.7616950842437674,
|
| 906 |
+
"eval_stsbenchmark_test_spearman": 0.7443131859049517,
|
| 907 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.76515246805641,
|
| 908 |
+
"step": 22500
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"epoch": 1.48,
|
| 912 |
+
"learning_rate": 5.391266851255051e-06,
|
| 913 |
+
"loss": 0.0001,
|
| 914 |
+
"step": 23125
|
| 915 |
+
},
|
| 916 |
+
{
|
| 917 |
+
"epoch": 1.48,
|
| 918 |
+
"eval_avg_sts": 0.7388169479527533,
|
| 919 |
+
"eval_avg_test": 0.7353151766024301,
|
| 920 |
+
"eval_avg_train_and_dev": 0.7396876104946917,
|
| 921 |
+
"eval_runtime": 20.7811,
|
| 922 |
+
"eval_sickrelatedness_spearman": 0.7059146816791875,
|
| 923 |
+
"eval_sickrelatedness_test_spearman": 0.7009398671844066,
|
| 924 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7109578919652116,
|
| 925 |
+
"eval_sts12_spearman": 0.6337385819368979,
|
| 926 |
+
"eval_sts13_spearman": 0.8077158631122653,
|
| 927 |
+
"eval_sts14_spearman": 0.7041881431485422,
|
| 928 |
+
"eval_sts15_spearman": 0.7897395308324451,
|
| 929 |
+
"eval_sts16_spearman": 0.7659376148408716,
|
| 930 |
+
"eval_stsbenchmark_spearman": 0.7644842201190634,
|
| 931 |
+
"eval_stsbenchmark_test_spearman": 0.7449466351615813,
|
| 932 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7684173290241716,
|
| 933 |
+
"step": 23125
|
| 934 |
+
},
|
| 935 |
+
{
|
| 936 |
+
"epoch": 1.52,
|
| 937 |
+
"learning_rate": 5.249648222958559e-06,
|
| 938 |
+
"loss": 0.0001,
|
| 939 |
+
"step": 23750
|
| 940 |
+
},
|
| 941 |
+
{
|
| 942 |
+
"epoch": 1.52,
|
| 943 |
+
"eval_avg_sts": 0.737637406431803,
|
| 944 |
+
"eval_avg_test": 0.7340377292492831,
|
| 945 |
+
"eval_avg_train_and_dev": 0.7393379830475855,
|
| 946 |
+
"eval_runtime": 20.8525,
|
| 947 |
+
"eval_sickrelatedness_spearman": 0.7079463598817589,
|
| 948 |
+
"eval_sickrelatedness_test_spearman": 0.7029335485554905,
|
| 949 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7129341890695166,
|
| 950 |
+
"eval_sts12_spearman": 0.6292541229208708,
|
| 951 |
+
"eval_sts13_spearman": 0.806837478246872,
|
| 952 |
+
"eval_sts14_spearman": 0.7027686303180887,
|
| 953 |
+
"eval_sts15_spearman": 0.7886320383460999,
|
| 954 |
+
"eval_sts16_spearman": 0.7662443034073775,
|
| 955 |
+
"eval_stsbenchmark_spearman": 0.7617789119015519,
|
| 956 |
+
"eval_stsbenchmark_test_spearman": 0.7415939829501823,
|
| 957 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7657417770256544,
|
| 958 |
+
"step": 23750
|
| 959 |
+
},
|
| 960 |
+
{
|
| 961 |
+
"epoch": 1.56,
|
| 962 |
+
"learning_rate": 5.107802641732104e-06,
|
| 963 |
+
"loss": 0.0001,
|
| 964 |
+
"step": 24375
|
| 965 |
+
},
|
| 966 |
+
{
|
| 967 |
+
"epoch": 1.56,
|
| 968 |
+
"eval_avg_sts": 0.7400432752757417,
|
| 969 |
+
"eval_avg_test": 0.7365741296789751,
|
| 970 |
+
"eval_avg_train_and_dev": 0.7403337201879336,
|
| 971 |
+
"eval_runtime": 21.0229,
|
| 972 |
+
"eval_sickrelatedness_spearman": 0.706139895856204,
|
| 973 |
+
"eval_sickrelatedness_test_spearman": 0.7011363476050757,
|
| 974 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7111677563573823,
|
| 975 |
+
"eval_sts12_spearman": 0.6393107034863543,
|
| 976 |
+
"eval_sts13_spearman": 0.8060305176934567,
|
| 977 |
+
"eval_sts14_spearman": 0.707469456143251,
|
| 978 |
+
"eval_sts15_spearman": 0.7884280597687867,
|
| 979 |
+
"eval_sts16_spearman": 0.7672135350116814,
|
| 980 |
+
"eval_stsbenchmark_spearman": 0.7657107589704576,
|
| 981 |
+
"eval_stsbenchmark_test_spearman": 0.7464302880442203,
|
| 982 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7694996840184851,
|
| 983 |
+
"step": 24375
|
| 984 |
+
},
|
| 985 |
+
{
|
| 986 |
+
"epoch": 1.6,
|
| 987 |
+
"learning_rate": 4.965957060505652e-06,
|
| 988 |
+
"loss": 0.0001,
|
| 989 |
+
"step": 25000
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"epoch": 1.6,
|
| 993 |
+
"eval_avg_sts": 0.7420831068144936,
|
| 994 |
+
"eval_avg_test": 0.7386039201251836,
|
| 995 |
+
"eval_avg_train_and_dev": 0.7419238844699756,
|
| 996 |
+
"eval_runtime": 20.7399,
|
| 997 |
+
"eval_sickrelatedness_spearman": 0.7086802746453288,
|
| 998 |
+
"eval_sickrelatedness_test_spearman": 0.703562769385137,
|
| 999 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7138105143408351,
|
| 1000 |
+
"eval_sts12_spearman": 0.643289234865721,
|
| 1001 |
+
"eval_sts13_spearman": 0.8077408521902738,
|
| 1002 |
+
"eval_sts14_spearman": 0.7094079968345621,
|
| 1003 |
+
"eval_sts15_spearman": 0.7910146281134862,
|
| 1004 |
+
"eval_sts16_spearman": 0.768300387542403,
|
| 1005 |
+
"eval_stsbenchmark_spearman": 0.7661483735096799,
|
| 1006 |
+
"eval_stsbenchmark_test_spearman": 0.7469115719447025,
|
| 1007 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.770037254599116,
|
| 1008 |
+
"step": 25000
|
| 1009 |
+
},
|
| 1010 |
+
{
|
| 1011 |
+
"epoch": 1.64,
|
| 1012 |
+
"learning_rate": 4.824111479279197e-06,
|
| 1013 |
+
"loss": 0.0001,
|
| 1014 |
+
"step": 25625
|
| 1015 |
+
},
|
| 1016 |
+
{
|
| 1017 |
+
"epoch": 1.64,
|
| 1018 |
+
"eval_avg_sts": 0.738553955422263,
|
| 1019 |
+
"eval_avg_test": 0.735270462011063,
|
| 1020 |
+
"eval_avg_train_and_dev": 0.74132969124333,
|
| 1021 |
+
"eval_runtime": 21.0644,
|
| 1022 |
+
"eval_sickrelatedness_spearman": 0.7107616379619491,
|
| 1023 |
+
"eval_sickrelatedness_test_spearman": 0.7068872624845219,
|
| 1024 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.714676339410284,
|
| 1025 |
+
"eval_sts12_spearman": 0.6304124246015519,
|
| 1026 |
+
"eval_sts13_spearman": 0.8044063778132966,
|
| 1027 |
+
"eval_sts14_spearman": 0.7046730335095999,
|
| 1028 |
+
"eval_sts15_spearman": 0.790737193065777,
|
| 1029 |
+
"eval_sts16_spearman": 0.764825533590085,
|
| 1030 |
+
"eval_stsbenchmark_spearman": 0.7640614874135817,
|
| 1031 |
+
"eval_stsbenchmark_test_spearman": 0.744951409012609,
|
| 1032 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.767983043076376,
|
| 1033 |
+
"step": 25625
|
| 1034 |
+
},
|
| 1035 |
+
{
|
| 1036 |
+
"epoch": 1.68,
|
| 1037 |
+
"learning_rate": 4.682492850982707e-06,
|
| 1038 |
+
"loss": 0.0001,
|
| 1039 |
+
"step": 26250
|
| 1040 |
+
},
|
| 1041 |
+
{
|
| 1042 |
+
"epoch": 1.68,
|
| 1043 |
+
"eval_avg_sts": 0.7337378478994732,
|
| 1044 |
+
"eval_avg_test": 0.7302481798812982,
|
| 1045 |
+
"eval_avg_train_and_dev": 0.7376796598903316,
|
| 1046 |
+
"eval_runtime": 20.8597,
|
| 1047 |
+
"eval_sickrelatedness_spearman": 0.7093236435437729,
|
| 1048 |
+
"eval_sickrelatedness_test_spearman": 0.7057306150582766,
|
| 1049 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7130156557412384,
|
| 1050 |
+
"eval_sts12_spearman": 0.617871248735412,
|
| 1051 |
+
"eval_sts13_spearman": 0.8025593367032153,
|
| 1052 |
+
"eval_sts14_spearman": 0.7006950877629325,
|
| 1053 |
+
"eval_sts15_spearman": 0.785270135862098,
|
| 1054 |
+
"eval_sts16_spearman": 0.7622963644366769,
|
| 1055 |
+
"eval_stsbenchmark_spearman": 0.7581491182522041,
|
| 1056 |
+
"eval_stsbenchmark_test_spearman": 0.7373144706104764,
|
| 1057 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7623436640394248,
|
| 1058 |
+
"step": 26250
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"epoch": 1.72,
|
| 1062 |
+
"learning_rate": 4.5406472697562535e-06,
|
| 1063 |
+
"loss": 0.0001,
|
| 1064 |
+
"step": 26875
|
| 1065 |
+
},
|
| 1066 |
+
{
|
| 1067 |
+
"epoch": 1.72,
|
| 1068 |
+
"eval_avg_sts": 0.7361414839625063,
|
| 1069 |
+
"eval_avg_test": 0.7327260474324794,
|
| 1070 |
+
"eval_avg_train_and_dev": 0.7408386426761701,
|
| 1071 |
+
"eval_runtime": 21.0874,
|
| 1072 |
+
"eval_sickrelatedness_spearman": 0.7130553203175825,
|
| 1073 |
+
"eval_sickrelatedness_test_spearman": 0.7085041459895807,
|
| 1074 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7176214412488239,
|
| 1075 |
+
"eval_sts12_spearman": 0.6206145648203354,
|
| 1076 |
+
"eval_sts13_spearman": 0.8054403345042884,
|
| 1077 |
+
"eval_sts14_spearman": 0.7043767904617289,
|
| 1078 |
+
"eval_sts15_spearman": 0.785161403859601,
|
| 1079 |
+
"eval_sts16_spearman": 0.7642432303964917,
|
| 1080 |
+
"eval_stsbenchmark_spearman": 0.7600987433775165,
|
| 1081 |
+
"eval_stsbenchmark_test_spearman": 0.7407418619953297,
|
| 1082 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7640558441035162,
|
| 1083 |
+
"step": 26875
|
| 1084 |
+
},
|
| 1085 |
+
{
|
| 1086 |
+
"epoch": 1.76,
|
| 1087 |
+
"learning_rate": 4.398801688529799e-06,
|
| 1088 |
+
"loss": 0.0001,
|
| 1089 |
+
"step": 27500
|
| 1090 |
+
},
|
| 1091 |
+
{
|
| 1092 |
+
"epoch": 1.76,
|
| 1093 |
+
"eval_avg_sts": 0.7385647317858953,
|
| 1094 |
+
"eval_avg_test": 0.7350557597229346,
|
| 1095 |
+
"eval_avg_train_and_dev": 0.7429292984249471,
|
| 1096 |
+
"eval_runtime": 20.7225,
|
| 1097 |
+
"eval_sickrelatedness_spearman": 0.7123203429942663,
|
| 1098 |
+
"eval_sickrelatedness_test_spearman": 0.7077860589944194,
|
| 1099 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7168975415894818,
|
| 1100 |
+
"eval_sts12_spearman": 0.6250000311651265,
|
| 1101 |
+
"eval_sts13_spearman": 0.8098932511821607,
|
| 1102 |
+
"eval_sts14_spearman": 0.7051633776714191,
|
| 1103 |
+
"eval_sts15_spearman": 0.787197723837324,
|
| 1104 |
+
"eval_sts16_spearman": 0.7654891602041461,
|
| 1105 |
+
"eval_stsbenchmark_spearman": 0.7648892354468243,
|
| 1106 |
+
"eval_stsbenchmark_test_spearman": 0.7448607150059462,
|
| 1107 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7689610552604125,
|
| 1108 |
+
"step": 27500
|
| 1109 |
+
},
|
| 1110 |
+
{
|
| 1111 |
+
"epoch": 1.8,
|
| 1112 |
+
"learning_rate": 4.2569561073033456e-06,
|
| 1113 |
+
"loss": 0.0001,
|
| 1114 |
+
"step": 28125
|
| 1115 |
+
},
|
| 1116 |
+
{
|
| 1117 |
+
"epoch": 1.8,
|
| 1118 |
+
"eval_avg_sts": 0.736228210862442,
|
| 1119 |
+
"eval_avg_test": 0.7323225373743962,
|
| 1120 |
+
"eval_avg_train_and_dev": 0.7422242610156418,
|
| 1121 |
+
"eval_runtime": 21.1674,
|
| 1122 |
+
"eval_sickrelatedness_spearman": 0.7117315495344994,
|
| 1123 |
+
"eval_sickrelatedness_test_spearman": 0.7062794242883943,
|
| 1124 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7173189540357382,
|
| 1125 |
+
"eval_sts12_spearman": 0.6181306965992698,
|
| 1126 |
+
"eval_sts13_spearman": 0.8063871249869112,
|
| 1127 |
+
"eval_sts14_spearman": 0.7026158508695518,
|
| 1128 |
+
"eval_sts15_spearman": 0.7867321226315551,
|
| 1129 |
+
"eval_sts16_spearman": 0.7652562131980097,
|
| 1130 |
+
"eval_stsbenchmark_spearman": 0.7627439182172969,
|
| 1131 |
+
"eval_stsbenchmark_test_spearman": 0.7408563290470811,
|
| 1132 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7671295679955453,
|
| 1133 |
+
"step": 28125
|
| 1134 |
+
},
|
| 1135 |
+
{
|
| 1136 |
+
"epoch": 1.84,
|
| 1137 |
+
"learning_rate": 4.115110526076892e-06,
|
| 1138 |
+
"loss": 0.0001,
|
| 1139 |
+
"step": 28750
|
| 1140 |
+
},
|
| 1141 |
+
{
|
| 1142 |
+
"epoch": 1.84,
|
| 1143 |
+
"eval_avg_sts": 0.7372002276835504,
|
| 1144 |
+
"eval_avg_test": 0.7334893393874011,
|
| 1145 |
+
"eval_avg_train_and_dev": 0.7419188871398015,
|
| 1146 |
+
"eval_runtime": 21.0224,
|
| 1147 |
+
"eval_sickrelatedness_spearman": 0.7115479327768666,
|
| 1148 |
+
"eval_sickrelatedness_test_spearman": 0.7061596518024446,
|
| 1149 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7169625169788694,
|
| 1150 |
+
"eval_sts12_spearman": 0.6201372334499261,
|
| 1151 |
+
"eval_sts13_spearman": 0.8081427416220239,
|
| 1152 |
+
"eval_sts14_spearman": 0.7058112992260926,
|
| 1153 |
+
"eval_sts15_spearman": 0.7871505208738109,
|
| 1154 |
+
"eval_sts16_spearman": 0.7648927226152069,
|
| 1155 |
+
"eval_stsbenchmark_spearman": 0.7627191432209255,
|
| 1156 |
+
"eval_stsbenchmark_test_spearman": 0.7421312061223029,
|
| 1157 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7668752573007337,
|
| 1158 |
+
"step": 28750
|
| 1159 |
+
},
|
| 1160 |
+
{
|
| 1161 |
+
"epoch": 1.88,
|
| 1162 |
+
"learning_rate": 3.9732649448504385e-06,
|
| 1163 |
+
"loss": 0.0002,
|
| 1164 |
+
"step": 29375
|
| 1165 |
+
},
|
| 1166 |
+
{
|
| 1167 |
+
"epoch": 1.88,
|
| 1168 |
+
"eval_avg_sts": 0.7391379223370436,
|
| 1169 |
+
"eval_avg_test": 0.7356189035069137,
|
| 1170 |
+
"eval_avg_train_and_dev": 0.7435335345244203,
|
| 1171 |
+
"eval_runtime": 20.9593,
|
| 1172 |
+
"eval_sickrelatedness_spearman": 0.7106110928678779,
|
| 1173 |
+
"eval_sickrelatedness_test_spearman": 0.7046551635771077,
|
| 1174 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7166877770505714,
|
| 1175 |
+
"eval_sts12_spearman": 0.6242789115119458,
|
| 1176 |
+
"eval_sts13_spearman": 0.8073323769451142,
|
| 1177 |
+
"eval_sts14_spearman": 0.7082426452989751,
|
| 1178 |
+
"eval_sts15_spearman": 0.7878887744476603,
|
| 1179 |
+
"eval_sts16_spearman": 0.7689858922579222,
|
| 1180 |
+
"eval_stsbenchmark_spearman": 0.7666257630298101,
|
| 1181 |
+
"eval_stsbenchmark_test_spearman": 0.7479485605096713,
|
| 1182 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7703792919982693,
|
| 1183 |
+
"step": 29375
|
| 1184 |
+
},
|
| 1185 |
+
{
|
| 1186 |
+
"epoch": 1.92,
|
| 1187 |
+
"learning_rate": 3.831419363623985e-06,
|
| 1188 |
+
"loss": 0.0001,
|
| 1189 |
+
"step": 30000
|
| 1190 |
+
},
|
| 1191 |
+
{
|
| 1192 |
+
"epoch": 1.92,
|
| 1193 |
+
"eval_avg_sts": 0.7379153295754598,
|
| 1194 |
+
"eval_avg_test": 0.7345327453977518,
|
| 1195 |
+
"eval_avg_train_and_dev": 0.7418020175283047,
|
| 1196 |
+
"eval_runtime": 20.9305,
|
| 1197 |
+
"eval_sickrelatedness_spearman": 0.7095180409050843,
|
| 1198 |
+
"eval_sickrelatedness_test_spearman": 0.7035887655223291,
|
| 1199 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7155211362349155,
|
| 1200 |
+
"eval_sts12_spearman": 0.6242234464567169,
|
| 1201 |
+
"eval_sts13_spearman": 0.8060611779379976,
|
| 1202 |
+
"eval_sts14_spearman": 0.704822019666627,
|
| 1203 |
+
"eval_sts15_spearman": 0.7862063807761319,
|
| 1204 |
+
"eval_sts16_spearman": 0.770166017643468,
|
| 1205 |
+
"eval_stsbenchmark_spearman": 0.7644102236421939,
|
| 1206 |
+
"eval_stsbenchmark_test_spearman": 0.7466614097809916,
|
| 1207 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7680828988216938,
|
| 1208 |
+
"step": 30000
|
| 1209 |
+
},
|
| 1210 |
+
{
|
| 1211 |
+
"epoch": 1.96,
|
| 1212 |
+
"learning_rate": 3.6898007353274935e-06,
|
| 1213 |
+
"loss": 0.0002,
|
| 1214 |
+
"step": 30625
|
| 1215 |
+
},
|
| 1216 |
+
{
|
| 1217 |
+
"epoch": 1.96,
|
| 1218 |
+
"eval_avg_sts": 0.7414417844482861,
|
| 1219 |
+
"eval_avg_test": 0.7377227000737309,
|
| 1220 |
+
"eval_avg_train_and_dev": 0.7426887168127401,
|
| 1221 |
+
"eval_runtime": 20.9682,
|
| 1222 |
+
"eval_sickrelatedness_spearman": 0.7083850749498312,
|
| 1223 |
+
"eval_sickrelatedness_test_spearman": 0.7020411335449008,
|
| 1224 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7147828329921959,
|
| 1225 |
+
"eval_sts12_spearman": 0.6398748755196823,
|
| 1226 |
+
"eval_sts13_spearman": 0.8091407530457168,
|
| 1227 |
+
"eval_sts14_spearman": 0.7086523177233283,
|
| 1228 |
+
"eval_sts15_spearman": 0.7887396499382829,
|
| 1229 |
+
"eval_sts16_spearman": 0.7687341048891995,
|
| 1230 |
+
"eval_stsbenchmark_spearman": 0.7665657150719614,
|
| 1231 |
+
"eval_stsbenchmark_test_spearman": 0.7468760658550058,
|
| 1232 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7705946006332842,
|
| 1233 |
+
"step": 30625
|
| 1234 |
+
},
|
| 1235 |
+
{
|
| 1236 |
+
"epoch": 2.0,
|
| 1237 |
+
"learning_rate": 3.54795515410104e-06,
|
| 1238 |
+
"loss": 0.0001,
|
| 1239 |
+
"step": 31250
|
| 1240 |
+
},
|
| 1241 |
+
{
|
| 1242 |
+
"epoch": 2.0,
|
| 1243 |
+
"eval_avg_sts": 0.740485017285834,
|
| 1244 |
+
"eval_avg_test": 0.7370491200682708,
|
| 1245 |
+
"eval_avg_train_and_dev": 0.7429999990678977,
|
| 1246 |
+
"eval_runtime": 20.7964,
|
| 1247 |
+
"eval_sickrelatedness_spearman": 0.7097765510038309,
|
| 1248 |
+
"eval_sickrelatedness_test_spearman": 0.7039997719017956,
|
| 1249 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.715634333306455,
|
| 1250 |
+
"eval_sts12_spearman": 0.6346257316776542,
|
| 1251 |
+
"eval_sts13_spearman": 0.8098227751570725,
|
| 1252 |
+
"eval_sts14_spearman": 0.7070680432539359,
|
| 1253 |
+
"eval_sts15_spearman": 0.7867814254070365,
|
| 1254 |
+
"eval_sts16_spearman": 0.7686156462120806,
|
| 1255 |
+
"eval_stsbenchmark_spearman": 0.766704948289228,
|
| 1256 |
+
"eval_stsbenchmark_test_spearman": 0.7484304468683206,
|
| 1257 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7703656648293403,
|
| 1258 |
+
"step": 31250
|
| 1259 |
+
},
|
| 1260 |
+
{
|
| 1261 |
+
"epoch": 2.04,
|
| 1262 |
+
"learning_rate": 3.4063365258045485e-06,
|
| 1263 |
+
"loss": 0.0002,
|
| 1264 |
+
"step": 31875
|
| 1265 |
+
},
|
| 1266 |
+
{
|
| 1267 |
+
"epoch": 2.04,
|
| 1268 |
+
"eval_avg_sts": 0.7410438701692595,
|
| 1269 |
+
"eval_avg_test": 0.7373422938815762,
|
| 1270 |
+
"eval_avg_train_and_dev": 0.7446059296458287,
|
| 1271 |
+
"eval_runtime": 20.9546,
|
| 1272 |
+
"eval_sickrelatedness_spearman": 0.7105618614355229,
|
| 1273 |
+
"eval_sickrelatedness_test_spearman": 0.7049301366548643,
|
| 1274 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7162726117635637,
|
| 1275 |
+
"eval_sts12_spearman": 0.6306599951032117,
|
| 1276 |
+
"eval_sts13_spearman": 0.807157187196233,
|
| 1277 |
+
"eval_sts14_spearman": 0.709921780686096,
|
| 1278 |
+
"eval_sts15_spearman": 0.7882494996283947,
|
| 1279 |
+
"eval_sts16_spearman": 0.7718392477227314,
|
| 1280 |
+
"eval_stsbenchmark_spearman": 0.7689175194126273,
|
| 1281 |
+
"eval_stsbenchmark_test_spearman": 0.7486382101795024,
|
| 1282 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7729392475280936,
|
| 1283 |
+
"step": 31875
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"epoch": 2.08,
|
| 1287 |
+
"learning_rate": 3.264490944578095e-06,
|
| 1288 |
+
"loss": 0.0002,
|
| 1289 |
+
"step": 32500
|
| 1290 |
+
},
|
| 1291 |
+
{
|
| 1292 |
+
"epoch": 2.08,
|
| 1293 |
+
"eval_avg_sts": 0.7406899068085288,
|
| 1294 |
+
"eval_avg_test": 0.737240537915657,
|
| 1295 |
+
"eval_avg_train_and_dev": 0.7446535659927934,
|
| 1296 |
+
"eval_runtime": 21.252,
|
| 1297 |
+
"eval_sickrelatedness_spearman": 0.7129782951458211,
|
| 1298 |
+
"eval_sickrelatedness_test_spearman": 0.707353500824854,
|
| 1299 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.718621422755247,
|
| 1300 |
+
"eval_sts12_spearman": 0.631196341321798,
|
| 1301 |
+
"eval_sts13_spearman": 0.8084161729926506,
|
| 1302 |
+
"eval_sts14_spearman": 0.7092196209507754,
|
| 1303 |
+
"eval_sts15_spearman": 0.7846086513351899,
|
| 1304 |
+
"eval_sts16_spearman": 0.7714800805004121,
|
| 1305 |
+
"eval_stsbenchmark_spearman": 0.7669301854130546,
|
| 1306 |
+
"eval_stsbenchmark_test_spearman": 0.748409397483919,
|
| 1307 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7706857092303399,
|
| 1308 |
+
"step": 32500
|
| 1309 |
+
},
|
| 1310 |
+
{
|
| 1311 |
+
"epoch": 2.12,
|
| 1312 |
+
"learning_rate": 3.1226453633516414e-06,
|
| 1313 |
+
"loss": 0.0001,
|
| 1314 |
+
"step": 33125
|
| 1315 |
+
},
|
| 1316 |
+
{
|
| 1317 |
+
"epoch": 2.12,
|
| 1318 |
+
"eval_avg_sts": 0.741908611558358,
|
| 1319 |
+
"eval_avg_test": 0.7385184801792626,
|
| 1320 |
+
"eval_avg_train_and_dev": 0.745921273906644,
|
| 1321 |
+
"eval_runtime": 20.8772,
|
| 1322 |
+
"eval_sickrelatedness_spearman": 0.712568145774737,
|
| 1323 |
+
"eval_sickrelatedness_test_spearman": 0.7071536357667555,
|
| 1324 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7180482272764128,
|
| 1325 |
+
"eval_sts12_spearman": 0.6276979028419505,
|
| 1326 |
+
"eval_sts13_spearman": 0.8109754368105846,
|
| 1327 |
+
"eval_sts14_spearman": 0.7100445056205082,
|
| 1328 |
+
"eval_sts15_spearman": 0.7892064188117265,
|
| 1329 |
+
"eval_sts16_spearman": 0.7727292323220317,
|
| 1330 |
+
"eval_stsbenchmark_spearman": 0.7701386387269664,
|
| 1331 |
+
"eval_stsbenchmark_test_spearman": 0.751822229081281,
|
| 1332 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7737943205368751,
|
| 1333 |
+
"step": 33125
|
| 1334 |
+
},
|
| 1335 |
+
{
|
| 1336 |
+
"epoch": 2.16,
|
| 1337 |
+
"learning_rate": 2.980799782125188e-06,
|
| 1338 |
+
"loss": 0.0002,
|
| 1339 |
+
"step": 33750
|
| 1340 |
+
},
|
| 1341 |
+
{
|
| 1342 |
+
"epoch": 2.16,
|
| 1343 |
+
"eval_avg_sts": 0.7387749148969024,
|
| 1344 |
+
"eval_avg_test": 0.7355275285778003,
|
| 1345 |
+
"eval_avg_train_and_dev": 0.7427705356417595,
|
| 1346 |
+
"eval_runtime": 21.9616,
|
| 1347 |
+
"eval_sickrelatedness_spearman": 0.7116362724362297,
|
| 1348 |
+
"eval_sickrelatedness_test_spearman": 0.7063792375154563,
|
| 1349 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7170108350529064,
|
| 1350 |
+
"eval_sts12_spearman": 0.6210675920148566,
|
| 1351 |
+
"eval_sts13_spearman": 0.809613648614316,
|
| 1352 |
+
"eval_sts14_spearman": 0.7068348739830796,
|
| 1353 |
+
"eval_sts15_spearman": 0.786987945092306,
|
| 1354 |
+
"eval_sts16_spearman": 0.7702320756630628,
|
| 1355 |
+
"eval_stsbenchmark_spearman": 0.765051996474467,
|
| 1356 |
+
"eval_stsbenchmark_test_spearman": 0.7475773271615254,
|
| 1357 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7685302362306128,
|
| 1358 |
+
"step": 33750
|
| 1359 |
+
},
|
| 1360 |
+
{
|
| 1361 |
+
"epoch": 2.2,
|
| 1362 |
+
"learning_rate": 2.8389542008987334e-06,
|
| 1363 |
+
"loss": 0.0001,
|
| 1364 |
+
"step": 34375
|
| 1365 |
+
},
|
| 1366 |
+
{
|
| 1367 |
+
"epoch": 2.2,
|
| 1368 |
+
"eval_avg_sts": 0.7372500677744379,
|
| 1369 |
+
"eval_avg_test": 0.7342484305550541,
|
| 1370 |
+
"eval_avg_train_and_dev": 0.7427227214188237,
|
| 1371 |
+
"eval_runtime": 21.98,
|
| 1372 |
+
"eval_sickrelatedness_spearman": 0.7124431266877929,
|
| 1373 |
+
"eval_sickrelatedness_test_spearman": 0.7076623305284827,
|
| 1374 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7173681019291466,
|
| 1375 |
+
"eval_sts12_spearman": 0.6203661142159484,
|
| 1376 |
+
"eval_sts13_spearman": 0.806796081405754,
|
| 1377 |
+
"eval_sts14_spearman": 0.7047306825878202,
|
| 1378 |
+
"eval_sts15_spearman": 0.7819332414401158,
|
| 1379 |
+
"eval_sts16_spearman": 0.7696611847093084,
|
| 1380 |
+
"eval_stsbenchmark_spearman": 0.764820043374326,
|
| 1381 |
+
"eval_stsbenchmark_test_spearman": 0.7485893789979496,
|
| 1382 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7680773409085008,
|
| 1383 |
+
"step": 34375
|
| 1384 |
+
},
|
| 1385 |
+
{
|
| 1386 |
+
"epoch": 2.24,
|
| 1387 |
+
"learning_rate": 2.69710861967228e-06,
|
| 1388 |
+
"loss": 0.0001,
|
| 1389 |
+
"step": 35000
|
| 1390 |
+
},
|
| 1391 |
+
{
|
| 1392 |
+
"epoch": 2.24,
|
| 1393 |
+
"eval_avg_sts": 0.7387507850343802,
|
| 1394 |
+
"eval_avg_test": 0.7355342211965702,
|
| 1395 |
+
"eval_avg_train_and_dev": 0.7445443682185053,
|
| 1396 |
+
"eval_runtime": 22.2171,
|
| 1397 |
+
"eval_sickrelatedness_spearman": 0.7125033872391564,
|
| 1398 |
+
"eval_sickrelatedness_test_spearman": 0.7073772850432812,
|
| 1399 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7177230874857234,
|
| 1400 |
+
"eval_sts12_spearman": 0.6247069431706919,
|
| 1401 |
+
"eval_sts13_spearman": 0.8046176559037094,
|
| 1402 |
+
"eval_sts14_spearman": 0.7071478374873903,
|
| 1403 |
+
"eval_sts15_spearman": 0.7836492394816458,
|
| 1404 |
+
"eval_sts16_spearman": 0.7707143442262178,
|
| 1405 |
+
"eval_stsbenchmark_spearman": 0.7679160877318494,
|
| 1406 |
+
"eval_stsbenchmark_test_spearman": 0.7505262430630557,
|
| 1407 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7713656489512873,
|
| 1408 |
+
"step": 35000
|
| 1409 |
+
},
|
| 1410 |
+
{
|
| 1411 |
+
"epoch": 2.28,
|
| 1412 |
+
"learning_rate": 2.5554899913757893e-06,
|
| 1413 |
+
"loss": 0.0001,
|
| 1414 |
+
"step": 35625
|
| 1415 |
+
},
|
| 1416 |
+
{
|
| 1417 |
+
"epoch": 2.28,
|
| 1418 |
+
"eval_avg_sts": 0.7377723938824481,
|
| 1419 |
+
"eval_avg_test": 0.7345737983826125,
|
| 1420 |
+
"eval_avg_train_and_dev": 0.7435797752746427,
|
| 1421 |
+
"eval_runtime": 22.0951,
|
| 1422 |
+
"eval_sickrelatedness_spearman": 0.7123552031305164,
|
| 1423 |
+
"eval_sickrelatedness_test_spearman": 0.7072368789839194,
|
| 1424 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7175580611123827,
|
| 1425 |
+
"eval_sts12_spearman": 0.6227870968906012,
|
| 1426 |
+
"eval_sts13_spearman": 0.8046259866606313,
|
| 1427 |
+
"eval_sts14_spearman": 0.7045085522245373,
|
| 1428 |
+
"eval_sts15_spearman": 0.7838932428340308,
|
| 1429 |
+
"eval_sts16_spearman": 0.7701283228301814,
|
| 1430 |
+
"eval_stsbenchmark_spearman": 0.7661083526066392,
|
| 1431 |
+
"eval_stsbenchmark_test_spearman": 0.7488365082543857,
|
| 1432 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7696014894369028,
|
| 1433 |
+
"step": 35625
|
| 1434 |
+
},
|
| 1435 |
+
{
|
| 1436 |
+
"epoch": 2.32,
|
| 1437 |
+
"learning_rate": 2.4136444101493353e-06,
|
| 1438 |
+
"loss": 0.0,
|
| 1439 |
+
"step": 36250
|
| 1440 |
+
},
|
| 1441 |
+
{
|
| 1442 |
+
"epoch": 2.32,
|
| 1443 |
+
"eval_avg_sts": 0.7386155577954275,
|
| 1444 |
+
"eval_avg_test": 0.7353702532577143,
|
| 1445 |
+
"eval_avg_train_and_dev": 0.7437646103034667,
|
| 1446 |
+
"eval_runtime": 20.9396,
|
| 1447 |
+
"eval_sickrelatedness_spearman": 0.7116453653128161,
|
| 1448 |
+
"eval_sickrelatedness_test_spearman": 0.7065116410259358,
|
| 1449 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7168516996463055,
|
| 1450 |
+
"eval_sts12_spearman": 0.6248182960777477,
|
| 1451 |
+
"eval_sts13_spearman": 0.8056761080437632,
|
| 1452 |
+
"eval_sts14_spearman": 0.7057631810447995,
|
| 1453 |
+
"eval_sts15_spearman": 0.7847325098633744,
|
| 1454 |
+
"eval_sts16_spearman": 0.7705122919119067,
|
| 1455 |
+
"eval_stsbenchmark_spearman": 0.7671611523135844,
|
| 1456 |
+
"eval_stsbenchmark_test_spearman": 0.7495777448364731,
|
| 1457 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7706775209606281,
|
| 1458 |
+
"step": 36250
|
| 1459 |
+
},
|
| 1460 |
+
{
|
| 1461 |
+
"epoch": 2.36,
|
| 1462 |
+
"learning_rate": 2.2717988289228813e-06,
|
| 1463 |
+
"loss": 0.0001,
|
| 1464 |
+
"step": 36875
|
| 1465 |
+
},
|
| 1466 |
+
{
|
| 1467 |
+
"epoch": 2.36,
|
| 1468 |
+
"eval_avg_sts": 0.740149603021757,
|
| 1469 |
+
"eval_avg_test": 0.7367482300057596,
|
| 1470 |
+
"eval_avg_train_and_dev": 0.7453363908257578,
|
| 1471 |
+
"eval_runtime": 20.671,
|
| 1472 |
+
"eval_sickrelatedness_spearman": 0.7126227371950188,
|
| 1473 |
+
"eval_sickrelatedness_test_spearman": 0.7077303136706309,
|
| 1474 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7176363979763565,
|
| 1475 |
+
"eval_sts12_spearman": 0.6274167410561244,
|
| 1476 |
+
"eval_sts13_spearman": 0.8069745763455181,
|
| 1477 |
+
"eval_sts14_spearman": 0.7076654105229765,
|
| 1478 |
+
"eval_sts15_spearman": 0.7863406428310873,
|
| 1479 |
+
"eval_sts16_spearman": 0.7707586061652102,
|
| 1480 |
+
"eval_stsbenchmark_spearman": 0.7692685070363641,
|
| 1481 |
+
"eval_stsbenchmark_test_spearman": 0.7503513194487701,
|
| 1482 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7730363836751589,
|
| 1483 |
+
"step": 36875
|
| 1484 |
+
},
|
| 1485 |
+
{
|
| 1486 |
+
"epoch": 2.4,
|
| 1487 |
+
"learning_rate": 2.129953247696428e-06,
|
| 1488 |
+
"loss": 0.0001,
|
| 1489 |
+
"step": 37500
|
| 1490 |
+
},
|
| 1491 |
+
{
|
| 1492 |
+
"epoch": 2.4,
|
| 1493 |
+
"eval_avg_sts": 0.7395131963250486,
|
| 1494 |
+
"eval_avg_test": 0.7361937524639587,
|
| 1495 |
+
"eval_avg_train_and_dev": 0.7453992073695682,
|
| 1496 |
+
"eval_runtime": 20.4199,
|
| 1497 |
+
"eval_sickrelatedness_spearman": 0.7127770844855997,
|
| 1498 |
+
"eval_sickrelatedness_test_spearman": 0.7078060121991877,
|
| 1499 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.717903112924754,
|
| 1500 |
+
"eval_sts12_spearman": 0.6274978324241501,
|
| 1501 |
+
"eval_sts13_spearman": 0.805585095341077,
|
| 1502 |
+
"eval_sts14_spearman": 0.7071221897597898,
|
| 1503 |
+
"eval_sts15_spearman": 0.7836729881285205,
|
| 1504 |
+
"eval_sts16_spearman": 0.7706690961987493,
|
| 1505 |
+
"eval_stsbenchmark_spearman": 0.7692680879374545,
|
| 1506 |
+
"eval_stsbenchmark_test_spearman": 0.7510030531962361,
|
| 1507 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7728953018143825,
|
| 1508 |
+
"step": 37500
|
| 1509 |
+
},
|
| 1510 |
+
{
|
| 1511 |
+
"epoch": 2.44,
|
| 1512 |
+
"learning_rate": 1.9883346193999368e-06,
|
| 1513 |
+
"loss": 0.0002,
|
| 1514 |
+
"step": 38125
|
| 1515 |
+
},
|
| 1516 |
+
{
|
| 1517 |
+
"epoch": 2.44,
|
| 1518 |
+
"eval_avg_sts": 0.7387624644292584,
|
| 1519 |
+
"eval_avg_test": 0.7352227757479233,
|
| 1520 |
+
"eval_avg_train_and_dev": 0.7425307943142997,
|
| 1521 |
+
"eval_runtime": 20.7947,
|
| 1522 |
+
"eval_sickrelatedness_spearman": 0.7102028117293744,
|
| 1523 |
+
"eval_sickrelatedness_test_spearman": 0.7047476872624795,
|
| 1524 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7157212183265218,
|
| 1525 |
+
"eval_sts12_spearman": 0.6320284509392846,
|
| 1526 |
+
"eval_sts13_spearman": 0.8033937882675368,
|
| 1527 |
+
"eval_sts14_spearman": 0.7054126300177038,
|
| 1528 |
+
"eval_sts15_spearman": 0.78530200196467,
|
| 1529 |
+
"eval_sts16_spearman": 0.7695255766427133,
|
| 1530 |
+
"eval_stsbenchmark_spearman": 0.7654719914435268,
|
| 1531 |
+
"eval_stsbenchmark_test_spearman": 0.7461492951410754,
|
| 1532 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7693403703020776,
|
| 1533 |
+
"step": 38125
|
| 1534 |
+
},
|
| 1535 |
+
{
|
| 1536 |
+
"epoch": 2.48,
|
| 1537 |
+
"learning_rate": 1.846715991103445e-06,
|
| 1538 |
+
"loss": 0.0001,
|
| 1539 |
+
"step": 38750
|
| 1540 |
+
},
|
| 1541 |
+
{
|
| 1542 |
+
"epoch": 2.48,
|
| 1543 |
+
"eval_avg_sts": 0.7374253093552523,
|
| 1544 |
+
"eval_avg_test": 0.7339404463063186,
|
| 1545 |
+
"eval_avg_train_and_dev": 0.7421398368400338,
|
| 1546 |
+
"eval_runtime": 20.4703,
|
| 1547 |
+
"eval_sickrelatedness_spearman": 0.7107438562185104,
|
| 1548 |
+
"eval_sickrelatedness_test_spearman": 0.7052514908771994,
|
| 1549 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7163253665937409,
|
| 1550 |
+
"eval_sts12_spearman": 0.6286712483831545,
|
| 1551 |
+
"eval_sts13_spearman": 0.8022301065346743,
|
| 1552 |
+
"eval_sts14_spearman": 0.7048882397124776,
|
| 1553 |
+
"eval_sts15_spearman": 0.7831405592800499,
|
| 1554 |
+
"eval_sts16_spearman": 0.7680903822018597,
|
| 1555 |
+
"eval_stsbenchmark_spearman": 0.7642127731560395,
|
| 1556 |
+
"eval_stsbenchmark_test_spearman": 0.7453110971548148,
|
| 1557 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7679543070863266,
|
| 1558 |
+
"step": 38750
|
| 1559 |
+
},
|
| 1560 |
+
{
|
| 1561 |
+
"epoch": 2.52,
|
| 1562 |
+
"learning_rate": 1.7048704098769918e-06,
|
| 1563 |
+
"loss": 0.0,
|
| 1564 |
+
"step": 39375
|
| 1565 |
+
},
|
| 1566 |
+
{
|
| 1567 |
+
"epoch": 2.52,
|
| 1568 |
+
"eval_avg_sts": 0.7368107941649409,
|
| 1569 |
+
"eval_avg_test": 0.7334652203074301,
|
| 1570 |
+
"eval_avg_train_and_dev": 0.741976935359566,
|
| 1571 |
+
"eval_runtime": 20.8182,
|
| 1572 |
+
"eval_sickrelatedness_spearman": 0.7104882673504456,
|
| 1573 |
+
"eval_sickrelatedness_test_spearman": 0.7051023764808654,
|
| 1574 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7159384852490909,
|
| 1575 |
+
"eval_sts12_spearman": 0.6268768757514214,
|
| 1576 |
+
"eval_sts13_spearman": 0.8029190740819949,
|
| 1577 |
+
"eval_sts14_spearman": 0.7022475030486164,
|
| 1578 |
+
"eval_sts15_spearman": 0.7826530446059501,
|
| 1579 |
+
"eval_sts16_spearman": 0.7681144502337414,
|
| 1580 |
+
"eval_stsbenchmark_spearman": 0.7643763440824164,
|
| 1581 |
+
"eval_stsbenchmark_test_spearman": 0.7463432179494219,
|
| 1582 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7680153854700411,
|
| 1583 |
+
"step": 39375
|
| 1584 |
+
},
|
| 1585 |
+
{
|
| 1586 |
+
"epoch": 2.56,
|
| 1587 |
+
"learning_rate": 1.5630248286505382e-06,
|
| 1588 |
+
"loss": 0.0001,
|
| 1589 |
+
"step": 40000
|
| 1590 |
+
},
|
| 1591 |
+
{
|
| 1592 |
+
"epoch": 2.56,
|
| 1593 |
+
"eval_avg_sts": 0.7381003659828818,
|
| 1594 |
+
"eval_avg_test": 0.7347244014162845,
|
| 1595 |
+
"eval_avg_train_and_dev": 0.7429134210348216,
|
| 1596 |
+
"eval_runtime": 20.3825,
|
| 1597 |
+
"eval_sickrelatedness_spearman": 0.710316206946611,
|
| 1598 |
+
"eval_sickrelatedness_test_spearman": 0.7046947189644344,
|
| 1599 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7160195342787384,
|
| 1600 |
+
"eval_sts12_spearman": 0.6302424800213983,
|
| 1601 |
+
"eval_sts13_spearman": 0.803247786425164,
|
| 1602 |
+
"eval_sts14_spearman": 0.7037417379983866,
|
| 1603 |
+
"eval_sts15_spearman": 0.7831307580280404,
|
| 1604 |
+
"eval_sts16_spearman": 0.7698214390074158,
|
| 1605 |
+
"eval_stsbenchmark_spearman": 0.7662021534531573,
|
| 1606 |
+
"eval_stsbenchmark_test_spearman": 0.7481918894691519,
|
| 1607 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7698073077909047,
|
| 1608 |
+
"step": 40000
|
| 1609 |
+
},
|
| 1610 |
+
{
|
| 1611 |
+
"epoch": 2.6,
|
| 1612 |
+
"learning_rate": 1.4211792474240842e-06,
|
| 1613 |
+
"loss": 0.0,
|
| 1614 |
+
"step": 40625
|
| 1615 |
+
},
|
| 1616 |
+
{
|
| 1617 |
+
"epoch": 2.6,
|
| 1618 |
+
"eval_avg_sts": 0.7384929628200189,
|
| 1619 |
+
"eval_avg_test": 0.7351061112883038,
|
| 1620 |
+
"eval_avg_train_and_dev": 0.7432958438543483,
|
| 1621 |
+
"eval_runtime": 20.8515,
|
| 1622 |
+
"eval_sickrelatedness_spearman": 0.7108589590676125,
|
| 1623 |
+
"eval_sickrelatedness_test_spearman": 0.7053178143923748,
|
| 1624 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7164885700394883,
|
| 1625 |
+
"eval_sts12_spearman": 0.6291888769664764,
|
| 1626 |
+
"eval_sts13_spearman": 0.8041919909911799,
|
| 1627 |
+
"eval_sts14_spearman": 0.7045984873525695,
|
| 1628 |
+
"eval_sts15_spearman": 0.7839088787514867,
|
| 1629 |
+
"eval_sts16_spearman": 0.7702165266681544,
|
| 1630 |
+
"eval_stsbenchmark_spearman": 0.766487019942652,
|
| 1631 |
+
"eval_stsbenchmark_test_spearman": 0.7483202038958853,
|
| 1632 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7701031176692084,
|
| 1633 |
+
"step": 40625
|
| 1634 |
+
},
|
| 1635 |
+
{
|
| 1636 |
+
"epoch": 2.64,
|
| 1637 |
+
"learning_rate": 1.2793336661976307e-06,
|
| 1638 |
+
"loss": 0.0002,
|
| 1639 |
+
"step": 41250
|
| 1640 |
+
},
|
| 1641 |
+
{
|
| 1642 |
+
"epoch": 2.64,
|
| 1643 |
+
"eval_avg_sts": 0.7371561236957961,
|
| 1644 |
+
"eval_avg_test": 0.7339566580465452,
|
| 1645 |
+
"eval_avg_train_and_dev": 0.7422583980781617,
|
| 1646 |
+
"eval_runtime": 21.234,
|
| 1647 |
+
"eval_sickrelatedness_spearman": 0.7099911047770248,
|
| 1648 |
+
"eval_sickrelatedness_test_spearman": 0.7044999663247513,
|
| 1649 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7155628833300955,
|
| 1650 |
+
"eval_sts12_spearman": 0.6254860905536815,
|
| 1651 |
+
"eval_sts13_spearman": 0.8017523519499081,
|
| 1652 |
+
"eval_sts14_spearman": 0.7028572928547238,
|
| 1653 |
+
"eval_sts15_spearman": 0.7845378195212266,
|
| 1654 |
+
"eval_sts16_spearman": 0.7699394553459527,
|
| 1655 |
+
"eval_stsbenchmark_spearman": 0.7655287508680542,
|
| 1656 |
+
"eval_stsbenchmark_test_spearman": 0.7486236297755721,
|
| 1657 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7689539128262279,
|
| 1658 |
+
"step": 41250
|
| 1659 |
+
},
|
| 1660 |
+
{
|
| 1661 |
+
"epoch": 2.68,
|
| 1662 |
+
"learning_rate": 1.1374880849711772e-06,
|
| 1663 |
+
"loss": 0.0001,
|
| 1664 |
+
"step": 41875
|
| 1665 |
+
},
|
| 1666 |
+
{
|
| 1667 |
+
"epoch": 2.68,
|
| 1668 |
+
"eval_avg_sts": 0.7370631114339077,
|
| 1669 |
+
"eval_avg_test": 0.7338136544671776,
|
| 1670 |
+
"eval_avg_train_and_dev": 0.7423301757389922,
|
| 1671 |
+
"eval_runtime": 21.6943,
|
| 1672 |
+
"eval_sickrelatedness_spearman": 0.7102428073706631,
|
| 1673 |
+
"eval_sickrelatedness_test_spearman": 0.7048236058355623,
|
| 1674 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7157821202465618,
|
| 1675 |
+
"eval_sts12_spearman": 0.6236458167934951,
|
| 1676 |
+
"eval_sts13_spearman": 0.8023700660147888,
|
| 1677 |
+
"eval_sts14_spearman": 0.703084563331531,
|
| 1678 |
+
"eval_sts15_spearman": 0.7846832363439064,
|
| 1679 |
+
"eval_sts16_spearman": 0.7700794957874387,
|
| 1680 |
+
"eval_stsbenchmark_spearman": 0.7653357943955308,
|
| 1681 |
+
"eval_stsbenchmark_test_spearman": 0.7480087971635208,
|
| 1682 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7688782312314226,
|
| 1683 |
+
"step": 41875
|
| 1684 |
+
},
|
| 1685 |
+
{
|
| 1686 |
+
"epoch": 2.72,
|
| 1687 |
+
"learning_rate": 9.956425037447234e-07,
|
| 1688 |
+
"loss": 0.0001,
|
| 1689 |
+
"step": 42500
|
| 1690 |
+
},
|
| 1691 |
+
{
|
| 1692 |
+
"epoch": 2.72,
|
| 1693 |
+
"eval_avg_sts": 0.7382956278861713,
|
| 1694 |
+
"eval_avg_test": 0.734991312963749,
|
| 1695 |
+
"eval_avg_train_and_dev": 0.7428571308285099,
|
| 1696 |
+
"eval_runtime": 21.0897,
|
| 1697 |
+
"eval_sickrelatedness_spearman": 0.7102825098008535,
|
| 1698 |
+
"eval_sickrelatedness_test_spearman": 0.7048535850875951,
|
| 1699 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7157899403058504,
|
| 1700 |
+
"eval_sts12_spearman": 0.6265558778138477,
|
| 1701 |
+
"eval_sts13_spearman": 0.8033181918156761,
|
| 1702 |
+
"eval_sts14_spearman": 0.7042972985487412,
|
| 1703 |
+
"eval_sts15_spearman": 0.7861678530544808,
|
| 1704 |
+
"eval_sts16_spearman": 0.7711254897645943,
|
| 1705 |
+
"eval_stsbenchmark_spearman": 0.7663221744050053,
|
| 1706 |
+
"eval_stsbenchmark_test_spearman": 0.7486208946613078,
|
| 1707 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7699243213511694,
|
| 1708 |
+
"step": 42500
|
| 1709 |
+
},
|
| 1710 |
+
{
|
| 1711 |
+
"epoch": 2.76,
|
| 1712 |
+
"learning_rate": 8.540238754482321e-07,
|
| 1713 |
+
"loss": 0.0001,
|
| 1714 |
+
"step": 43125
|
| 1715 |
+
},
|
| 1716 |
+
{
|
| 1717 |
+
"epoch": 2.76,
|
| 1718 |
+
"eval_avg_sts": 0.7377715580722582,
|
| 1719 |
+
"eval_avg_test": 0.7345108563336945,
|
| 1720 |
+
"eval_avg_train_and_dev": 0.7427287762752324,
|
| 1721 |
+
"eval_runtime": 22.1235,
|
| 1722 |
+
"eval_sickrelatedness_spearman": 0.7106771021682546,
|
| 1723 |
+
"eval_sickrelatedness_test_spearman": 0.705266686228856,
|
| 1724 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.716181272162837,
|
| 1725 |
+
"eval_sts12_spearman": 0.6239496815809444,
|
| 1726 |
+
"eval_sts13_spearman": 0.8039159573225746,
|
| 1727 |
+
"eval_sts14_spearman": 0.7040700656997784,
|
| 1728 |
+
"eval_sts15_spearman": 0.7851629218072443,
|
| 1729 |
+
"eval_sts16_spearman": 0.7708450368779131,
|
| 1730 |
+
"eval_stsbenchmark_spearman": 0.7657801410490988,
|
| 1731 |
+
"eval_stsbenchmark_test_spearman": 0.7483656448185506,
|
| 1732 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7692762803876276,
|
| 1733 |
+
"step": 43125
|
| 1734 |
+
},
|
| 1735 |
+
{
|
| 1736 |
+
"epoch": 2.8,
|
| 1737 |
+
"learning_rate": 7.121782942217785e-07,
|
| 1738 |
+
"loss": 0.0,
|
| 1739 |
+
"step": 43750
|
| 1740 |
+
},
|
| 1741 |
+
{
|
| 1742 |
+
"epoch": 2.8,
|
| 1743 |
+
"eval_avg_sts": 0.7376141103751033,
|
| 1744 |
+
"eval_avg_test": 0.7344015898023963,
|
| 1745 |
+
"eval_avg_train_and_dev": 0.742544402311891,
|
| 1746 |
+
"eval_runtime": 21.1096,
|
| 1747 |
+
"eval_sickrelatedness_spearman": 0.7104606827644014,
|
| 1748 |
+
"eval_sickrelatedness_test_spearman": 0.7050746381321955,
|
| 1749 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7159137410961507,
|
| 1750 |
+
"eval_sts12_spearman": 0.6239536269253912,
|
| 1751 |
+
"eval_sts13_spearman": 0.8036508752187724,
|
| 1752 |
+
"eval_sts14_spearman": 0.7037219130584743,
|
| 1753 |
+
"eval_sts15_spearman": 0.7847939731114079,
|
| 1754 |
+
"eval_sts16_spearman": 0.7709951067177447,
|
| 1755 |
+
"eval_stsbenchmark_spearman": 0.7657225948295306,
|
| 1756 |
+
"eval_stsbenchmark_test_spearman": 0.748620995452788,
|
| 1757 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7691750635276314,
|
| 1758 |
+
"step": 43750
|
| 1759 |
+
},
|
| 1760 |
+
{
|
| 1761 |
+
"epoch": 2.84,
|
| 1762 |
+
"learning_rate": 5.703327129953248e-07,
|
| 1763 |
+
"loss": 0.0001,
|
| 1764 |
+
"step": 44375
|
| 1765 |
+
},
|
| 1766 |
+
{
|
| 1767 |
+
"epoch": 2.84,
|
| 1768 |
+
"eval_avg_sts": 0.7375152940520443,
|
| 1769 |
+
"eval_avg_test": 0.7343389041124367,
|
| 1770 |
+
"eval_avg_train_and_dev": 0.7423452174385804,
|
| 1771 |
+
"eval_runtime": 22.0931,
|
| 1772 |
+
"eval_sickrelatedness_spearman": 0.7104542745458605,
|
| 1773 |
+
"eval_sickrelatedness_test_spearman": 0.705098130115833,
|
| 1774 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7158819129939784,
|
| 1775 |
+
"eval_sts12_spearman": 0.6246511970226402,
|
| 1776 |
+
"eval_sts13_spearman": 0.8034092134157865,
|
| 1777 |
+
"eval_sts14_spearman": 0.7037963451422589,
|
| 1778 |
+
"eval_sts15_spearman": 0.7839004357252624,
|
| 1779 |
+
"eval_sts16_spearman": 0.7710148359050526,
|
| 1780 |
+
"eval_stsbenchmark_spearman": 0.7653807566074485,
|
| 1781 |
+
"eval_stsbenchmark_test_spearman": 0.7485021714602236,
|
| 1782 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7688085218831824,
|
| 1783 |
+
"step": 44375
|
| 1784 |
+
},
|
| 1785 |
+
{
|
| 1786 |
+
"epoch": 2.88,
|
| 1787 |
+
"learning_rate": 4.2848713176887115e-07,
|
| 1788 |
+
"loss": 0.0,
|
| 1789 |
+
"step": 45000
|
| 1790 |
+
},
|
| 1791 |
+
{
|
| 1792 |
+
"epoch": 2.88,
|
| 1793 |
+
"eval_avg_sts": 0.7373122545485458,
|
| 1794 |
+
"eval_avg_test": 0.7341422495112051,
|
| 1795 |
+
"eval_avg_train_and_dev": 0.7421676000848536,
|
| 1796 |
+
"eval_runtime": 21.0333,
|
| 1797 |
+
"eval_sickrelatedness_spearman": 0.7104022241403064,
|
| 1798 |
+
"eval_sickrelatedness_test_spearman": 0.7050898274710157,
|
| 1799 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7157835004453044,
|
| 1800 |
+
"eval_sts12_spearman": 0.6238646284703343,
|
| 1801 |
+
"eval_sts13_spearman": 0.8032636806389909,
|
| 1802 |
+
"eval_sts14_spearman": 0.7035512511045346,
|
| 1803 |
+
"eval_sts15_spearman": 0.7839647615882837,
|
| 1804 |
+
"eval_sts16_spearman": 0.7710192044322828,
|
| 1805 |
+
"eval_stsbenchmark_spearman": 0.7651200314650877,
|
| 1806 |
+
"eval_stsbenchmark_test_spearman": 0.7482423928729935,
|
| 1807 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.768551699724403,
|
| 1808 |
+
"step": 45000
|
| 1809 |
+
},
|
| 1810 |
+
{
|
| 1811 |
+
"epoch": 2.92,
|
| 1812 |
+
"learning_rate": 2.8664155054241755e-07,
|
| 1813 |
+
"loss": 0.0001,
|
| 1814 |
+
"step": 45625
|
| 1815 |
+
},
|
| 1816 |
+
{
|
| 1817 |
+
"epoch": 2.92,
|
| 1818 |
+
"eval_avg_sts": 0.7376021690547506,
|
| 1819 |
+
"eval_avg_test": 0.7344364718587105,
|
| 1820 |
+
"eval_avg_train_and_dev": 0.7420987124013947,
|
| 1821 |
+
"eval_runtime": 20.9052,
|
| 1822 |
+
"eval_sickrelatedness_spearman": 0.7100355064557035,
|
| 1823 |
+
"eval_sickrelatedness_test_spearman": 0.7047251628468761,
|
| 1824 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7154114262576732,
|
| 1825 |
+
"eval_sts12_spearman": 0.6245606765009919,
|
| 1826 |
+
"eval_sts13_spearman": 0.8038598745242214,
|
| 1827 |
+
"eval_sts14_spearman": 0.7038508600251897,
|
| 1828 |
+
"eval_sts15_spearman": 0.7847361674768141,
|
| 1829 |
+
"eval_sts16_spearman": 0.770818571348578,
|
| 1830 |
+
"eval_stsbenchmark_spearman": 0.765353527051755,
|
| 1831 |
+
"eval_stsbenchmark_test_spearman": 0.7485039902883023,
|
| 1832 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7687859985451162,
|
| 1833 |
+
"step": 45625
|
| 1834 |
+
},
|
| 1835 |
+
{
|
| 1836 |
+
"epoch": 2.96,
|
| 1837 |
+
"learning_rate": 1.447959693159639e-07,
|
| 1838 |
+
"loss": 0.0001,
|
| 1839 |
+
"step": 46250
|
| 1840 |
+
},
|
| 1841 |
+
{
|
| 1842 |
+
"epoch": 2.96,
|
| 1843 |
+
"eval_avg_sts": 0.737426533458678,
|
| 1844 |
+
"eval_avg_test": 0.7342698102173869,
|
| 1845 |
+
"eval_avg_train_and_dev": 0.7419274116141954,
|
| 1846 |
+
"eval_runtime": 20.2997,
|
| 1847 |
+
"eval_sickrelatedness_spearman": 0.7099669691358331,
|
| 1848 |
+
"eval_sickrelatedness_test_spearman": 0.7046739918719009,
|
| 1849 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7153334695992425,
|
| 1850 |
+
"eval_sts12_spearman": 0.6238235463591946,
|
| 1851 |
+
"eval_sts13_spearman": 0.8041066380633766,
|
| 1852 |
+
"eval_sts14_spearman": 0.7036262868879434,
|
| 1853 |
+
"eval_sts15_spearman": 0.7846043548200616,
|
| 1854 |
+
"eval_sts16_spearman": 0.770752451815878,
|
| 1855 |
+
"eval_stsbenchmark_spearman": 0.7651054871284582,
|
| 1856 |
+
"eval_stsbenchmark_test_spearman": 0.7483014017033541,
|
| 1857 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7685213536291483,
|
| 1858 |
+
"step": 46250
|
| 1859 |
+
},
|
| 1860 |
+
{
|
| 1861 |
+
"epoch": 3.0,
|
| 1862 |
+
"learning_rate": 2.9503880895102356e-09,
|
| 1863 |
+
"loss": 0.0001,
|
| 1864 |
+
"step": 46875
|
| 1865 |
+
},
|
| 1866 |
+
{
|
| 1867 |
+
"epoch": 3.0,
|
| 1868 |
+
"eval_avg_sts": 0.7374977429972559,
|
| 1869 |
+
"eval_avg_test": 0.7343576643139398,
|
| 1870 |
+
"eval_avg_train_and_dev": 0.7419344676757816,
|
| 1871 |
+
"eval_runtime": 20.8011,
|
| 1872 |
+
"eval_sickrelatedness_spearman": 0.7099435140291807,
|
| 1873 |
+
"eval_sickrelatedness_test_spearman": 0.7046738346759807,
|
| 1874 |
+
"eval_sickrelatedness_train_and_dev_spearman": 0.7152876184590635,
|
| 1875 |
+
"eval_sts12_spearman": 0.6240417964491191,
|
| 1876 |
+
"eval_sts13_spearman": 0.8041867824864875,
|
| 1877 |
+
"eval_sts14_spearman": 0.703700075594037,
|
| 1878 |
+
"eval_sts15_spearman": 0.7846174552488144,
|
| 1879 |
+
"eval_sts16_spearman": 0.7708088556462593,
|
| 1880 |
+
"eval_stsbenchmark_spearman": 0.7651857215268932,
|
| 1881 |
+
"eval_stsbenchmark_test_spearman": 0.7484748500968808,
|
| 1882 |
+
"eval_stsbenchmark_train_and_dev_spearman": 0.7685813168924998,
|
| 1883 |
+
"step": 46875
|
| 1884 |
+
},
|
| 1885 |
+
{
|
| 1886 |
+
"epoch": 3.0,
|
| 1887 |
+
"step": 46875,
|
| 1888 |
+
"total_flos": 0.0,
|
| 1889 |
+
"train_loss": 0.00036472132327159247,
|
| 1890 |
+
"train_runtime": 8039.9662,
|
| 1891 |
+
"train_samples_per_second": 373.136,
|
| 1892 |
+
"train_steps_per_second": 5.83
|
| 1893 |
+
}
|
| 1894 |
+
],
|
| 1895 |
+
"max_steps": 46875,
|
| 1896 |
+
"num_train_epochs": 3,
|
| 1897 |
+
"total_flos": 0.0,
|
| 1898 |
+
"trial_name": null,
|
| 1899 |
+
"trial_params": null
|
| 1900 |
+
}
|