| from typing import List, Dict, Tuple |
| import re |
|
|
| from data.base_env import DynamicEnv |
| from data.utils.retrieval_utils import Retriever |
|
|
| class TriviaQAEnv(DynamicEnv): |
| |
| def __init__(self, configs: Dict): |
| super().__init__(configs) |
| self.explorer = Retriever() |
|
|
| def set_env(self, task_config: Dict) -> None: |
| if task_config.get('answer') is None: |
| raise ValueError('Please provide the answer for the task') |
| if task_config.get("prompt") is None: |
| raise ValueError('Please provide the prompt for the task') |
|
|
| self.task_config = task_config |
| |
| self._reset() |
|
|
| from data.triviaqa.builder import TRIVIAQA_SYSTEM_PROMPT |
| return TRIVIAQA_SYSTEM_PROMPT, task_config["prompt"] |
| |
| def _reset(self): |
| self.done = False |
| self.reward = 0.0 |
|
|
| def step(self, action: str) -> Tuple[str, float, bool]: |
| action = self.preprocess_action(action) |
| action_type, action_content = self._process_action(action) |
| observation = None |
| |
| if action_type == "search": |
| try: |
| observation = self.explorer.batch_search([action_content])[0] |
| except Exception as e: |
| observation = f'Cannot find corresponding pages.' |
| self.done = False |
| self.reward = 0.0 |
|
|
| elif action_type == "answer": |
| observation = "" |
| self.done = True |
| self.reward = 1.0 if self._check_answer(action_content, self.task_config["answer"]) else 0.0 |
| else: |
| observation = "\nMy previous action is invalid. \ |
| If I want to search, I should put the query between <search> and </search>. \ |
| If I want to give the final answer, I should put the answer between <answer> and </answer>. Let me try again.\n" |
| self.done = False |
| self.reward = 0.0 |
|
|
| return observation, self.reward, self.done |
| |
| @classmethod |
| def preprocess_action(cls, action: str) -> str: |
| if "</search>" in action: |
| return action.split("</search>", 1)[0] + "</search>" |
| elif "</answer>" in action: |
| return action.split("</answer>", 1)[0] + "</answer>" |
| else: |
| return action |
| |
| @classmethod |
| def _process_action(cls, action: str): |
| action = action.strip() |
|
|
| if "<search>" in action and "</search>" in action: |
| start = action.index("<search>") + len("<search>") |
| end = action.index("</search>") |
| content = action[start:end].strip().split("\n", 1)[0].strip() |
| return "search", content |
|
|
| if "<answer>" in action and "</answer>" in action: |
| start = action.index("<answer>") + len("<answer>") |
| end = action.index("</answer>") |
| content = action[start:end].strip().split("\n", 1)[0].strip() |
| return "answer", content |
|
|
| return "think", action |
| |
| def _check_answer(self, answer: str, ground_truth: List[str]): |
| answer = answer.lower() |
| for gt in ground_truth: |
| if gt.lower() in answer: |
| return True |
| |
| return False |
| |
| def feedback(self) -> float: |
| return self.reward |
|
|
| @classmethod |
| def compute_reward(cls, completions: List[str], envs: List['TriviaQAEnv'], **kwargs) -> List[float]: |
| scores = [] |
| for completion, env in zip(completions, envs): |
| solution = env.task_config['answer'] |
|
|
| matches = re.findall(r"<answer>(.*?)</answer>", completion, re.DOTALL) |
| |
| if not matches: |
| scores.append(0.0) |
| continue |
|
|
| extracted = matches[-1].strip() |
|
|
| correct = False |
| for s in solution: |
| if s.lower() in extracted.lower(): |
| correct = True |
| break |
|
|
| if correct: |
| scores.append(1.0) |
| else: |
| scores.append(0.5) |
|
|
| return scores |
|
|
|
|
