Upload grpo_training.py with huggingface_hub

grpo_training.py

@@ -68,12 +68,12 @@ config = GRPOConfig(
 )
 
 # Define reward function for GRPO
-def preference_reward_func(samples):
+def preference_reward_func(inputs, prompts, completions, completion_ids_list):
     """Simple reward function based on response length preference"""
     rewards = []
-    for sample in samples:
+    for completion in completions:
         # Prefer shorter, more concise responses (addressing verbosity issue)
-        response_length = len(sample["response"].split())
+        response_length = len(completion.split())
         # Reward shorter responses (up to a reasonable length)
         if response_length < 50:
             reward = 1.0