RL parising: Greedy match outermost </reasoning> tag

rm stop strings

Author: hengtaoguo

src/maxtext/configs/post_train/rl.yml

@@ -152,7 +152,7 @@ max_num_seqs: null
 # If True, enables asynchronous scheduling in vLLM for faster generation
 async_scheduling: True
 # stop generation when any of these strings is generated
-stop_strings: [</answer>]
+stop_strings: null
 
 # ====== Checkpoint Configuration ======
 enable_checkpointing: True

src/maxtext/trainers/post_train/rl/utils_rl.py

@@ -106,7 +106,7 @@ def get_match_format_regex(tmvp_config):
   match_format = re.compile(
       (
           r"^[\s]{0,}"
-          rf"{tmvp_config.reasoning_start_token}.+?{tmvp_config.reasoning_end_token}.*?"
+          rf"{tmvp_config.reasoning_start_token}.+{tmvp_config.reasoning_end_token}.*?"
           rf"{tmvp_config.solution_start_token}(.+?){tmvp_config.solution_end_token}"
       ),
       flags=re.MULTILINE | re.DOTALL,