Commit
•
1b0fdc6
1
Parent(s):
b32f244
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
datasets:
|
| 3 |
+
- klue
|
| 4 |
+
- wikipedia
|
| 5 |
+
language:
|
| 6 |
+
- ko
|
| 7 |
+
metrics:
|
| 8 |
+
- accuracy
|
| 9 |
+
training_args:
|
| 10 |
+
- num_train_epochs=5,
|
| 11 |
+
- per_device_train_batch_size=16,
|
| 12 |
+
- per_device_eval_batch_size=16,
|
| 13 |
+
- prediction_loss_only=False,
|
| 14 |
+
- learning_rate=5e-5,
|
| 15 |
+
- logging_strategy='steps',
|
| 16 |
+
- logging_steps=100,
|
| 17 |
+
- save_steps=1000,
|
| 18 |
+
- eval_steps=1000,
|
| 19 |
+
- save_strategy="steps",
|
| 20 |
+
- evaluation_strategy="steps",
|
| 21 |
+
- load_best_model_at_end=True,
|
| 22 |
+
- metric_for_best_model="masked_accuracy",
|
| 23 |
+
- greater_is_better=True,
|
| 24 |
+
- seed=42,
|
| 25 |
+
- warmup_steps=5000,
|
| 26 |
+
info:
|
| 27 |
+
- MLM (15%) from the checkpoint of klue/roberta-large
|
| 28 |
+
- LineByLineTextDataset (block_size 384)
|
| 29 |
+
- PLM for ODQA task based-on Wikipedia questions
|
| 30 |
+
- Accuracy (for [MASK]) = 0.7066 (CE loss 1.388)
|
| 31 |
+
- v2 is trained with smaller learning rate and more epochs
|
| 32 |
+
---
|