Update README.md
Browse files
README.md
CHANGED
|
@@ -228,6 +228,24 @@ The following hyperparameters were used during training:
|
|
| 228 |
- data type: TF32
|
| 229 |
- num_epochs: 2
|
| 230 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 231 |
### Framework versions
|
| 232 |
|
| 233 |
- Transformers 4.22.0
|
|
|
|
| 228 |
- data type: TF32
|
| 229 |
- num_epochs: 2
|
| 230 |
|
| 231 |
+
#### Epochs 7 & 8
|
| 232 |
+
|
| 233 |
+
- epochs 5 & 6 were trained with 12288 tokens input
|
| 234 |
+
- this fixes that with 2 epochs at 16384 tokens input
|
| 235 |
+
|
| 236 |
+
The following hyperparameters were used during training:
|
| 237 |
+
- learning_rate: 0.0004
|
| 238 |
+
- train_batch_size: 4
|
| 239 |
+
- eval_batch_size: 1
|
| 240 |
+
- seed: 42
|
| 241 |
+
- distributed_type: multi-GPU
|
| 242 |
+
- gradient_accumulation_steps: 16
|
| 243 |
+
- total_train_batch_size: 64
|
| 244 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 245 |
+
- lr_scheduler_type: cosine
|
| 246 |
+
- lr_scheduler_warmup_ratio: 0.03
|
| 247 |
+
- num_epochs: 2
|
| 248 |
+
|
| 249 |
### Framework versions
|
| 250 |
|
| 251 |
- Transformers 4.22.0
|