Fix code style with make precommit (#4119)

scripts/generate_zen_dataset.py

@@ -338,7 +338,7 @@ def main(test_size, push_to_hub, repo_id):
             "Namespaces are one",
             "Although practicality sometimes beats purity,",
         ],
-        "completions":[
+        "completions": [
             [", let me think...", " ugly."],
             [", of course,", " implicit.", " because clarity matters."],
             ["... let's keep it basic,", " complex."],
@@ -350,7 +350,7 @@ def main(test_size, push_to_hub, repo_id):
             [" some theoretical elegance,", " purity."],
             [" silently,", " unless explicitly silenced."],
             [" the temptation to guess."],
-            [" way to do it,"," but sometimes it's not obvious.", " especially when there's more than one possibility."],
+            [" way to do it,", " but sometimes it's not obvious.", " especially when there's more than one possibility."],
             [" clear at first,", " it will eventually emerge."],
             [" later."],
             [" problematic fixes."],

tests/test_sft_trainer.py

@@ -1411,10 +1411,9 @@ class SFTTrainerTester(TrlTestCase):
     def test_peft_model_with_quantization(self):
         """SFTTrainer should not freeze layers of existing PeftModel.
 
-        This test simulates a realistic QLoRA scenario where a quantized base model
-        is first converted to a PeftModel, then passed to SFTTrainer. The issue was
-        that prepare_model_for_kbit_training would freeze all parameters including
-        the LoRA adapters, making training impossible.
+        This test simulates a realistic QLoRA scenario where a quantized base model is first converted to a PeftModel,
+        then passed to SFTTrainer. The issue was that prepare_model_for_kbit_training would freeze all parameters
+        including the LoRA adapters, making training impossible.
         """
         # Get the base model
         model_id = "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5"

trl/trainer/ppo_trainer.py

@@ -271,7 +271,7 @@ class PPOTrainer(Trainer):
         )  # note that we are calling `self.lr_scheduler.step()` manually only at the batch level
 
         #########
-        ### trainer specifics
+        # trainer specifics
         #########
         default_callbacks = DEFAULT_CALLBACKS + get_reporting_integration_callbacks(self.args.report_to)
         self.callbacks = default_callbacks if callbacks is None else default_callbacks + callbacks
@@ -303,7 +303,7 @@ class PPOTrainer(Trainer):
             self.model.add_model_tags(self._tag_names)
 
         #########
-        ### setup dataloader
+        # setup dataloader
         #########
         self.dataloader = DataLoader(
             self.train_dataset,