Update train.py

train.py

@@ -7,10 +7,10 @@ from transformers import AutoTokenizer, LlamaConfig, LlamaForCausalLM, TrainingA
 from datasets import load_dataset, DatasetDict, Dataset
 from tokenizers import ByteLevelBPETokenizer
 
-BATCH_SIZE = 96
-EPOCHS = 10
+BATCH_SIZE = 4
+EPOCHS = 2
 LEARNING_RATE = 2e-4
-FACTOR = 512
+FACTOR = 22 * 69
 MAX_SEQ_LENGTH = 128
 VOCAB_SIZE = 32000
 INPUT_DATASET = "HuggingFaceTB/smollm-corpus"
@@ -24,9 +24,9 @@ PUSH_TO_HUB = True
 
 def load_data():
     pretrain = load_dataset(INPUT_DATASET, "cosmopedia-v2", split="train", streaming=True)
-    pretrain = Dataset.from_generator(lambda: pretrain.take(int(3e+5)))
+    pretrain = Dataset.from_generator(lambda: pretrain.take(int(3e+4)))
     instruct = load_dataset(INSTRUCT_DATASET, split="train", streaming=True)
-    instruct = Dataset.from_generator(lambda: instruct.take(int(5e+5)))
+    instruct = Dataset.from_generator(lambda: instruct.take(int(5e+4)))
     dataset_dict = DatasetDict({
         'pretrain': pretrain,
         'instruct': instruct