diff --git a/pyproject.toml b/pyproject.toml
index b538578fd5..ae46d7aaf5 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -37,11 +37,11 @@ triton = [
 ]
 
 huggingface = [
-    "unsloth_zoo>=2025.7.2",
+    "unsloth_zoo>=2025.7.4",
     "packaging",
     "tyro",
-    "transformers>=4.51.3,!=4.47.0,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3",
-    "datasets>=3.4.1",
+    "transformers>=4.51.3,!=4.47.0,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0",
+    "datasets>=3.4.1,<4.0.0",
     "sentencepiece>=0.2.0",
     "tqdm",
     "psutil",
@@ -381,11 +381,11 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3",
 ]
 colab-new = [
-    "unsloth_zoo>=2025.7.2",
+    "unsloth_zoo>=2025.7.4",
     "packaging",
     "tyro",
-    "transformers>=4.51.3,!=4.47.0,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3",
-    "datasets>=3.4.1",
+    "transformers>=4.51.3,!=4.47.0,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0",
+    "datasets>=3.4.1,<4.0.0",
     "sentencepiece>=0.2.0",
     "tqdm",
     "psutil",
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 5076c422c8..c2520a66b5 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2025.7.2"
+__version__ = "2025.7.3"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 002e4d1863..4d5b4d7a4d 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -176,7 +176,7 @@ def grpo_trainer__prepare_inputs(function_name, function):
     import re
     # This matches the function signature, decorators and any comments immediately following
     pattern = r"(\s*@profiling_decorator\s*\n\s*def _prepare_inputs\s*\([^\)]*\)\s*(->\s*[^:]+)?\s*:\s*\n(?:[ ]*#[^\n]*\n)*)"
-    
+
     match = re.search(pattern, function)
     insert = (
         "        if hasattr(self, 'llm'):\n"
@@ -196,7 +196,7 @@ def grpo_trainer__prepare_inputs(function_name, function):
             rest_of_function,
             flags=re.DOTALL | re.MULTILINE
         )
-        
+
         # We also need to remove the old wake up call from the beginning of the function
         # since it's injected before the comments.
         header_and_comments = re.sub(
@@ -373,7 +373,11 @@ def grpo_trainer_compute_loss(function_name, function):
         _input_ids = input_ids
         _logits_to_keep = logits_to_keep
 
-        get_logps_func = lambda model, input_ids, attention_mask, logits_to_keep, batch_size=None, compute_entropy=False: self._get_per_token_logps(model, input_ids, attention_mask, logits_to_keep, batch_size) if hasattr(self, "_get_per_token_logps") else self._get_per_token_logps_and_entropies(model, input_ids, attention_mask, logits_to_keep, batch_size, compute_entropy)['logps']
+        get_logps_func = \
+            lambda model, input_ids, attention_mask, logits_to_keep, batch_size=None, compute_entropy=False: \
+            self._get_per_token_logps(model, input_ids, attention_mask, logits_to_keep) \
+            if hasattr(self, "_get_per_token_logps") else \
+            self._get_per_token_logps_and_entropies(model, input_ids, attention_mask, logits_to_keep, batch_size, compute_entropy)['logps']
 
         per_token_logps = get_logps_func(model, input_ids, attention_mask, logits_to_keep)