mlm-2

README.md

@@ -16,7 +16,7 @@ should probably proofread and complete it, then remove this comment. -->
 
 This model is a fine-tuned version of [aubmindlab/bert-base-arabertv02-twitter](https://huggingface.co/aubmindlab/bert-base-arabertv02-twitter) on the nuha-dataset dataset.
 It achieves the following results on the evaluation set:
-- Loss: 6.7276
+- Loss: 4.7143
 
 ## Model description
 
@@ -35,23 +35,29 @@ More information needed
 ### Training hyperparameters
 
 The following hyperparameters were used during training:
-- learning_rate: 1e-05
-- train_batch_size: 32
+- learning_rate: 5e-05
+- train_batch_size: 16
 - eval_batch_size: 32
 - seed: 42
-- gradient_accumulation_steps: 2
+- gradient_accumulation_steps: 4
 - total_train_batch_size: 64
 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
 - lr_scheduler_type: constant
 - lr_scheduler_warmup_steps: 1000.0
-- num_epochs: 1
+- num_epochs: 5
 - label_smoothing_factor: 0.1
 
 ### Training results
 
 | Training Loss | Epoch | Step | Validation Loss |
 |:-------------:|:-----:|:----:|:---------------:|
-| 7.028         | 0.64  | 500  | 6.7276          |
+| 5.6551        | 0.64  | 500  | 5.3465          |
+| 5.2489        | 1.28  | 1000 | 5.1498          |
+| 5.1004        | 1.91  | 1500 | 4.9168          |
+| 4.9289        | 2.55  | 2000 | 4.9202          |
+| 4.9081        | 3.19  | 2500 | 4.8600          |
+| 4.7641        | 3.83  | 3000 | 4.7688          |
+| 4.6543        | 4.47  | 3500 | 4.7143          |
 
 
 ### Framework versions

config.json

@@ -15,7 +15,7 @@
   "max_position_embeddings": 512,
   "model_type": "bert",
   "num_attention_heads": 12,
-  "num_hidden_layers": 4,
+  "num_hidden_layers": 8,
   "pad_token_id": 0,
   "position_embedding_type": "absolute",
   "torch_dtype": "float32",

config.toml

@@ -1,5 +1,5 @@
 [experiment]
-name = "mlm-1"
+name = "mlm-2"
 type = "mlm"
 
 
@@ -13,25 +13,26 @@ augment_ratio = 0.0
 [model]
 pretrained_model_name_or_path = "aubmindlab/bert-base-arabertv02-twitter"
 revision = "main"
-#hidden_dropout_prob = 0
-#attention_probs_dropout_prob = 0
-#classifier_dropout = 0
-num_hidden_layers = 4
+#hidden_dropout_prob = 0.2
+#attention_probs_dropout_prob = 0.2
+#classifier_dropout = 0.2
+num_hidden_layers = 8 
 #num_attention_heads = 12
 #hidden_size = 768
 #intermediate_size= null
 
 
 [training]
-num_train_epochs = 1
+num_train_epochs = 5
 warmup_steps = 1e3
 lr_scheduler_type = "constant"
-learning_rate = 1e-5
-per_device_train_batch_size = 32 
-per_device_eval_batch_size = 32 
-gradient_accumulation_steps = 2
+learning_rate = 5e-5
+per_device_train_batch_size = 16 
+per_device_eval_batch_size = 32
+gradient_accumulation_steps = 4
 weight_decay = 0.01
 label_smoothing_factor = 0.1
 weighted_loss = false
-early_stopping_patience = 5
-early_stopping_threshold = 0.005
+resample_data = false 
+early_stopping_patience = 0
+early_stopping_threshold = 0

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2390bcc18655c3b705f4bc3a92dfa3e334a5545ec9a2ea7203aad44f4dda0972
-size 314248719
+oid sha256:77ba656ca0a468c5254acd0f409bf4736a1cbf0140d9064fd0e21c59cbe176a3
+size 427676597

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5816c68c135c2d5cb8ef366758286e5d2acaca037248be513e6bdfcb64fb6c8c
+oid sha256:14b7b9084803fcab4baf4a0bb822aa03dd0f81f6311c38ae7f3411489a2dd872
 size 4027