fix on-policy data order

Author: yumeng5

on_policy_data_gen/decode.py

@@ -30,7 +30,7 @@
 
 train_dataset= load_dataset(data_dir, split='train_prefs')
 
-prompts = list(set(train_dataset['prompt']))
+prompts = sorted(list(set(train_dataset['prompt'])))
 
 conversations = [tokenizer.apply_chat_template([{'role': 'user', 'content': prompt}], tokenize=False, add_generation_prompt=True) for prompt in prompts]