diff --git a/dev/data/fineweb.py b/dev/data/fineweb.py
index 5109d31..12e1eb7 100644
--- a/dev/data/fineweb.py
+++ b/dev/data/fineweb.py
@@ -67,8 +67,8 @@ elif args.type =="edu":
 def tokenize_llama(doc):
     # tokenizes a single document and returns a numpy array of uint32 tokens
     tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B")
-    encode = lambda s: tokenizer.encode(s, add_special_tokens=False, verbose=False)
-    eot = 128000 # BOS; hardcoding because I don't know how to get it properly TODO
+    encode = lambda s: tokenizer.encode(s, add_special_tokens=False, verbose=False, split_special_tokens=True)
+    eot = tokenizer.encode('')[0] # by default the tokenizer adds the EOT token (128000)
     tokens = [eot] # the special <|endoftext|> token delimits all documents
     tokens.extend(encode(doc["text"]))
     tokens_np = np.array(tokens)
diff --git a/dev/data/tinyshakespeare.py b/dev/data/tinyshakespeare.py
index e5ba0d9..758a9a4 100644
--- a/dev/data/tinyshakespeare.py
+++ b/dev/data/tinyshakespeare.py
@@ -51,8 +51,8 @@ def tokenize(model_desc):
         eot = enc._special_tokens['<|endoftext|>'] # end of text token
     elif model_desc == "llama-3":
         tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B")
-        encode = lambda s: tokenizer.encode(s, add_special_tokens=False, verbose=False)
-        eot = 128000 # BOS; hardcoding because I don't know how to get it properly TODO
+        encode = lambda s: tokenizer.encode(s, add_special_tokens=False, verbose=False, split_special_tokens=True)
+        eot = tokenizer.encode('')[0] # by default the tokenizer adds the EOT token (128000)
     else:
         raise ValueError(f"unknown model descriptor {model_desc}")
     data_filename = os.path.join(DATA_CACHE_DIR, "tiny_shakespeare.txt")
diff --git a/dev/data/tinystories.py b/dev/data/tinystories.py
index aafd2cc..915c321 100644
--- a/dev/data/tinystories.py
+++ b/dev/data/tinystories.py
@@ -77,8 +77,8 @@ def process_shard(shard_index, shard_filename, model_desc):
         eot = enc._special_tokens['<|endoftext|>'] # end of text token
     elif model_desc == "llama-3":
         tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B")
-        encode = lambda s: tokenizer.encode(s, add_special_tokens=False, verbose=False)
-        eot = 128000 # BOS; hardcoding because I don't know how to get it properly TODO
+        encode = lambda s: tokenizer.encode(s, add_special_tokens=False, verbose=False, split_special_tokens=True)
+        eot = tokenizer.encode('')[0] # by default the tokenizer adds the EOT token (128000)
     else:
         raise ValueError(f"unknown model descriptor {model_desc}")