Commit d75b64b8 authored by jameskrw's avatar jameskrw Committed by YaningGao
Browse files

use strict match

parent a53c10cc

scripts/exps/format/navigation/free_think/run.sh

+2 −1
Original line number Diff line number Diff line
@@ -26,7 +26,7 @@ python3 -m vagen.trainer.main_ppo \ 
    algorithm.high_level_gamma=0.95 \

    data.train_files=data/$EXPERIMENT_NAME/train.parquet \

    data.val_files=data/$EXPERIMENT_NAME/test.parquet \

    data.train_batch_size=64 \

    data.train_batch_size=128 \

    data.val_batch_size=8 \

    data.max_prompt_length=1024 \

    data.max_response_length=150 \
@@ -83,5 +83,6 @@ python3 -m vagen.trainer.main_ppo \ 
    rollout_manager.n_trajectory=1 \

    rollout_manager.use_service=True \

    rollout_manager.timeout=240 \

    +rollout_manager.mini_batch_size=64 \

    rollout_manager.base_url="http://localhost:5000" \

    2>&1 | tee $EXPERIMENT_NAME.log

vagen/env/utils/parse_utils.py

+24 −24
Original line number Diff line number Diff line
@@ -15,16 +15,16 @@ def parse_freethink(response: str, special_token_list=None, action_sep=',', max_ 
    - format_correct: whether the response strictly follows the expected format

    """

    #Pattern to check for content strictly in the format <think>...</think><answer>...</answer>

    # strict_pattern = r'^\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*$'

    # strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    strict_pattern = r'^\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*$'

    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    

    

    # Pattern to extract content from think and answer tags

    extraction_pattern = r'<think>(.*?)</think>\s*<answer>(.*?)</answer>'

    match = re.search(extraction_pattern, response, re.DOTALL)

    format_correct = match is not None

    format_correct = strict_match is not None

    

    if not match:

    if not strict_match:

        think_content, action_content, actions = "", "", []

    else:

        think_content, action_content = match.group(1), match.group(2)
@@ -60,16 +60,16 @@ def parse_no_think(response: str, special_token_list=None, action_sep=',', max_a 
    - format_correct: whether the response strictly follows the expected format

    """

    # Pattern to check for content strictly in the format <answer>...</answer>

    # strict_pattern = r'^\s*<answer>(.*?)</answer>\s*$'

    # strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    # format_correct = strict_match is not None

    strict_pattern = r'^\s*<answer>(.*?)</answer>\s*$'

    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    format_correct = strict_match is not None

    

    # Pattern to extract content from answer tag

    extraction_pattern = r'<answer>(.*?)</answer>'

    match = re.search(extraction_pattern, response, re.DOTALL)

    format_correct = match is not None

    #format_correct = match is not None

    

    if not match:

    if not strict_match:

        think_content, action_content, actions = "", "", []

    else:

        action_content = match.group(1)
@@ -106,16 +106,16 @@ def parse_grounding(response: str, special_token_list=None, action_sep=',', max_ 
    - format_correct: whether the response strictly follows the expected format

    """

    # Pattern to check for content strictly in the expected format

    # strict_pattern = r'^\s*<current_state>(.*?)</current_state>\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*$'

    # strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    # format_correct = strict_match is not None

    strict_pattern = r'^\s*<current_state>(.*?)</current_state>\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*$'

    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    format_correct = strict_match is not None

    

    # Pattern to extract content from tags

    extraction_pattern = r'<current_state>(.*?)</current_state>\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>'

    match = re.search(extraction_pattern, response, re.DOTALL)

    format_correct = match is not None

    #format_correct = match is not None

    

    if not match:

    if not strict_match:

        current_state_content, think_content, action_content, actions = "", "", "", []

    else:

        current_state_content, think_content, action_content = match.group(1), match.group(2), match.group(3)
@@ -156,16 +156,16 @@ def parse_worldmodeling(response: str, special_token_list=None, action_sep=',', 
    - format_correct: whether the response strictly follows the expected format

    """

    # Pattern to check for content strictly in the expected format

    # strict_pattern = r'^\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*<next_state>(.*?)</next_state>\s*$'

    # strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    # format_correct = strict_match is not None

    strict_pattern = r'^\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*<next_state>(.*?)</next_state>\s*$'

    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    format_correct = strict_match is not None

    

    # Pattern to extract content from tags

    extraction_pattern = r'<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*<next_state>(.*?)</next_state>'

    match = re.search(extraction_pattern, response, re.DOTALL)

    format_correct = match is not None

    #format_correct = match is not None

    

    if not match:

    if not strict_match:

        think_content, action_content, next_state_content, actions = "", "", "", []

    else:

        think_content, action_content, next_state_content = match.group(1), match.group(2), match.group(3)
@@ -207,16 +207,16 @@ def parse_grounding_worldmodeling(response: str, special_token_list=None, action 
    - format_correct: whether the response strictly follows the expected format

    """

    # Pattern to check for content strictly in the expected format

    # strict_pattern = r'^\s*<current_state>(.*?)</current_state>\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*<next_state>(.*?)</next_state>\s*$'

    # strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    # format_correct = strict_match is not None

    strict_pattern = r'^\s*<current_state>(.*?)</current_state>\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*<next_state>(.*?)</next_state>\s*$'

    strict_match = re.match(strict_pattern, response.strip(), re.DOTALL)

    format_correct = strict_match is not None

    

    # Pattern to extract content from tags

    extraction_pattern = r'<current_state>(.*?)</current_state>\s*<think>(.*?)</think>\s*<answer>(.*?)</answer>\s*<next_state>(.*?)</next_state>'

    match = re.search(extraction_pattern, response, re.DOTALL)

    format_correct = match is not None

    #format_correct = match is not None

    

    if not match:

    if not strict_match:

        current_state_content, think_content, action_content, next_state_content, actions = "", "", "", "", []

    else:

        current_state_content, think_content, action_content, next_state_content = match.group(1), match.group(2), match.group(3), match.group(4)

vagen/trainer/ppo/ray_trainer.py

+1 −1
Original line number Diff line number Diff line
@@ -999,7 +999,7 @@ class RayPPOTrainer(object): 
                                                    prefix=logging_prefix)

        metrics.update(global_balance_stats)



    def _process_in_mini_batches(batch, rollout_manager, mini_batch_size):

    def _process_in_mini_batches(self,batch, rollout_manager, mini_batch_size):

        """

        Process the batch in mini-batches.

CRA Git | Maintained and supported by SUSTech CRA and CCSE