Update README.md
Browse files
README.md
CHANGED
|
@@ -57,10 +57,25 @@ _reliability_. While many of its answers are factually accurate, some are not. T
|
|
| 57 |
|
| 58 |
### Training hyperparameters
|
| 59 |
|
| 60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
The following hyperparameters were used during QA tuning:
|
| 63 |
- learning_rate: 2e-06
|
|
|
|
| 64 |
- train_batch_size: 4
|
| 65 |
- eval_batch_size: 4
|
| 66 |
- seed: 702
|
|
@@ -71,4 +86,5 @@ The following hyperparameters were used during QA tuning:
|
|
| 71 |
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 72 |
- lr_scheduler_type: linear
|
| 73 |
- lr_scheduler_warmup_steps: 100
|
| 74 |
-
- num_epochs: 2.0
|
|
|
|
|
|
| 57 |
|
| 58 |
### Training hyperparameters
|
| 59 |
|
| 60 |
+
The following hyperparameters were used during continued pretraining:
|
| 61 |
+
- learning_rate: 1e-05
|
| 62 |
+
- max_grad_norm: 3.0
|
| 63 |
+
- train_batch_size: 4
|
| 64 |
+
- eval_batch_size: 4
|
| 65 |
+
- seed: 701
|
| 66 |
+
- distributed_type: multi-GPU
|
| 67 |
+
- num_devices: 4
|
| 68 |
+
- total_train_batch_size: 16
|
| 69 |
+
- total_eval_batch_size: 16
|
| 70 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 71 |
+
- lr_scheduler_type: cosine
|
| 72 |
+
- lr_scheduler_warmup_steps: 100
|
| 73 |
+
- num_epochs: 3.0
|
| 74 |
+
- weight_decay: 1e-04
|
| 75 |
|
| 76 |
The following hyperparameters were used during QA tuning:
|
| 77 |
- learning_rate: 2e-06
|
| 78 |
+
- max_grad_norm: 3.0
|
| 79 |
- train_batch_size: 4
|
| 80 |
- eval_batch_size: 4
|
| 81 |
- seed: 702
|
|
|
|
| 86 |
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 87 |
- lr_scheduler_type: linear
|
| 88 |
- lr_scheduler_warmup_steps: 100
|
| 89 |
+
- num_epochs: 2.0
|
| 90 |
+
- weight_decay: 0.0
|