diff --git a/configs/experimental/roberta/roberta.oscar.nofilter.wechsel.largedict.sophia.json b/configs/experimental/roberta/roberta.oscar.nofilter.wechsel.largedict.sophia.json
new file mode 100644
index 0000000..daa1446
--- /dev/null
+++ b/configs/experimental/roberta/roberta.oscar.nofilter.wechsel.largedict.sophia.json
@@ -0,0 +1,30 @@
+{
+    "output_dir": "exps/roberta.oscar.nofilter.wechsel.largedict.sophia",
+    "model_type": "roberta",
+    "config_name": "exps/roberta.oscar.nofilter.wechsel.largedict.sophia",
+    "tokenizer_name": "exps/roberta.oscar.nofilter.wechsel.largedict.sophia",
+    "train_file": "data/oscar.nofilter",
+    "max_seq_length": 512,
+    "do_train": true,
+    "do_eval": true,
+    "validation_file": "data/bruk_valid_data.txt",
+    "weight_decay": 1e-1,
+    "per_device_train_batch_size": 24,
+    "per_device_eval_batch_size": 24,
+    "gradient_accumulation_steps": 4,
+    "learning_rate": 2e-4,
+    "evaluation_strategy": "steps",
+    "warmup_steps": 25000,
+    "max_steps": 250000,
+    "eval_steps": 500,
+    "save_steps": 12500,
+    "logging_steps": 500,
+    "overwrite_output_dir": true,
+    "num_train_epochs": 0,
+    "adam_beta1": 0.965,
+    "adam_beta2": 0.99,
+    "adam_epsilon": 1e-6,
+    "preprocessing_num_workers": 48,
+    "fp16": true,
+    "model_name_or_path": "exps/roberta.oscar.nofilter.wechsel.largedict.sophia"
+}
\ No newline at end of file
diff --git a/run_mlm.py b/run_mlm.py
index b9af68a..e7b3040 100644
--- a/run_mlm.py
+++ b/run_mlm.py
@@ -180,6 +180,11 @@ class DataTrainingArguments:
         metadata={"help": "The name of the project to which the training run will belong on Weights & Biases."}   
     )
 
+    optimizer: Optional[str] = field(
+        default="adam",
+        metadata={"help": "The optimizer to use for training."}
+    )
+
 
 class DataCollatorForLanguageModeling(transformers.DataCollatorForLanguageModeling):
     def torch_mask_tokens(self, inputs, special_tokens_mask = None):
@@ -503,17 +508,40 @@ def compute_metrics(eval_preds):
         wandb.config.update(data_args)
         wandb.save(__file__, policy="now")
 
-    # Initialize our Trainer
-    trainer = Trainer(
-        model=model,
-        args=training_args,
-        train_dataset=train_dataset if training_args.do_train else None,
-        eval_dataset=eval_dataset if training_args.do_eval else None,
-        tokenizer=tokenizer,
-        data_collator=data_collator,
-        compute_metrics=compute_metrics if training_args.do_eval else None,
-        preprocess_logits_for_metrics=preprocess_logits_for_metrics if training_args.do_eval else None,
-    )
+    if data_args.optimizer == "adam":
+        # Initialize our Trainer
+        trainer = Trainer(
+            model=model,
+            args=training_args,
+            train_dataset=train_dataset if training_args.do_train else None,
+            eval_dataset=eval_dataset if training_args.do_eval else None,
+            tokenizer=tokenizer,
+            data_collator=data_collator,
+            compute_metrics=compute_metrics if training_args.do_eval else None,
+            preprocess_logits_for_metrics=preprocess_logits_for_metrics if training_args.do_eval else None,
+        )
+    elif data_args.optimizer == "sophia":
+        from optimizers.sophia import SophiaG
+        optimizer = SophiaG(
+            filter(lambda p: p.requires_grad, model.parameters()),
+            lr=training_args.learning_rate,
+            weight_decay=training_args.weight_decay,
+            betas=(training_args.adam_beta1, training_args.adam_beta2),
+        )
+
+        trainer = Trainer(
+            model=model,
+            args=training_args,
+            train_dataset=train_dataset if training_args.do_train else None,
+            eval_dataset=eval_dataset if training_args.do_eval else None,
+            tokenizer=tokenizer,
+            data_collator=data_collator,
+            compute_metrics=compute_metrics if training_args.do_eval else None,
+            preprocess_logits_for_metrics=preprocess_logits_for_metrics if training_args.do_eval else None,
+            optimizers=(optimizer, None),
+        )
+    else:
+        raise ValueError("Optimizer not supported")
 
     # Training
     if training_args.do_train: