src/sal/search/diverse_verifier_tree_search.py (4 lines):
	- line 87: # TODO: set the augmented template from a file
	- line 108: # rarely ~1/1000 the model will generate few beams than expected. #TODO: investigate why
	- line 115: # scoring and chose best generation per beam TODO: add option for selection across beams within the same prompt
	- line 189: # TODO: construct and store the tree


src/sal/models/reward_models.py (3 lines):
	- line 118: # TODO: tokenize each batch independently so there is less padding and faster inference
	- line 132: # stripped_output_scores = [] TODO: strip out the reward for previous steps
	- line 187: # TODO: add the system prompt like we did for math shepard?


src/sal/utils/qwen_math_parser.py (2 lines):
	- line 380: # TODO: SFT models
	- line 488: # TODO check multiple choice


src/sal/search/best_of_n.py (1 line):
	- line 35: # TODO: set the augmented template from a file