Update train.py

train.py

@@ -8,9 +8,9 @@ from datasets import load_dataset, DatasetDict, Dataset
 from tokenizers import ByteLevelBPETokenizer
 
 MAX_SEQ_LENGTH = 2048
-BATCH_SIZE = 1
+BATCH_SIZE = 4
 EPOCHS = 2
-LEARNING_RATE = 1e-4
+LEARNING_RATE = 2e-2
 FACTOR = 1024
 VOCAB_SIZE = 3200
 INPUT_DATASET = "HuggingFaceTB/smollm-corpus"
@@ -18,7 +18,7 @@ INSTRUCT_DATASET = "nroggendorff/openhermes"
 OUTPUT_REPO = "smallama"
 FP16 = True
 WARMUP_STEPS = 20
-DECAY = 2e-2
+DECAY = 1e-4
 GRADIENT_ACCUMULATION_STEPS = 1
 CLIPPING = 1.0
 PUSH_TO_HUB = True