plancraft 0.3.6__py3-none-any.whl → 0.3.8__py3-none-any.whl

plancraft/environment/actions.py

@@ -70,7 +70,7 @@ class ActionHandlerBase(abc.ABC):
         raise NotImplementedError()
 
     @abc.abstractmethod
-    def match(self, generated_text: str):
+    def match(self, generated_text: str, **kwargs) -> Optional[BaseModel | str]:
         """
         Match the generated text to the action/tool
         """
@@ -204,7 +204,7 @@ class MoveActionHandler(ActionHandlerBase):
     def action_name(self) -> str:
         return "move"
 
-    def match(self, generated_text: str) -> Optional[MoveAction | str]:
+    def match(self, generated_text: str, **kwargs) -> Optional[MoveAction | str]:
         """
         Parse the raw model response to a MoveAction
         """
@@ -238,7 +238,7 @@ class SmeltActionHandler(ActionHandlerBase):
     def action_name(self) -> str:
         return "smelt"
 
-    def match(self, generated_text: str) -> Optional[SmeltAction | str]:
+    def match(self, generated_text: str, **kwargs) -> Optional[SmeltAction | str]:
         """
         Parse the raw model response to a SmeltAction
         """
@@ -272,7 +272,7 @@ class ImpossibleActionHandler(ActionHandlerBase):
     def action_name(self) -> str:
         return "impossible"
 
-    def match(self, generated_text) -> Optional[StopAction]:
+    def match(self, generated_text, **kwargs) -> Optional[StopAction]:
         """
         Parse the raw model response to a StopAction
         """
@@ -296,7 +296,7 @@ class ThinkActionHandler(ActionHandlerBase):
     def action_name(self) -> str:
         return "think"
 
-    def match(self, generated_text) -> Optional[str]:
+    def match(self, generated_text, **kwargs) -> Optional[str]:
         """
         Parse the raw model response to a ThinkAction
         """

plancraft/environment/search.py

@@ -42,7 +42,7 @@ class GoldSearchActionHandler(ActionHandlerBase):
     def action_name(self) -> str:
         return "search"
 
-    def match(self, generated_text) -> Optional[str]:
+    def match(self, generated_text, **kwargs) -> Optional[str]:
         """
         Parse the raw model response to a SearchAction
         """

plancraft/evaluator.py

@@ -142,13 +142,15 @@ class Evaluator:
                 return True
         return False
 
-    def parse_raw_model_response(self, generated_text: str):
+    def parse_raw_model_response(self, generated_text: str, observation=None) -> str:
         """
         Given a message and set of action handlers, parse the content to return the action
         or a message if the action is not valid/requires message response
         """
         for handler in self.actions:
-            match_output = handler.match(generated_text)
+            match_output = handler.match(
+                generated_text, observation=observation, history=self.history
+            )
             if match_output:
                 return match_output
         action_names = [handler.action_name for handler in self.actions]
@@ -192,7 +194,6 @@ class Evaluator:
         """Given the loaded model and an example from Plancraft
         run the episode until success or termination."""
         success = False
-        num_non_env_actions = 0
         self.reset(example)
         action = None
 
@@ -207,24 +208,18 @@ class Evaluator:
                 success = example.impossible
                 break
             # action is external tool then it is str
-            # limit the number of consecutive non-env actions to 3
-            elif isinstance(action, str) and num_non_env_actions < 3:
-                observation = {"message": action}
-                num_non_env_actions += 1
+            if isinstance(action, str):
+                observation = self.environment.step()
+                observation["target"] = example.target
+                observation["message"] = action
             # action is environment action
             else:
-                if isinstance(action, str):
-                    observation = self.environment.step()
-                else:
-                    observation = self.environment.step(action)
-
+                observation = self.environment.step(action)
                 # convert inventory observation to text message
                 observation["target"] = example.target
                 observation["message"] = self.convert_observation_to_message(
                     observation
                 )
-                num_non_env_actions = 0
-
                 # check if the episode is done
                 success = self.check_done(observation["inventory"], example.target)
             # exit if success
@@ -242,7 +237,7 @@ class Evaluator:
             # add message to history
             self.history.add_message_to_history(content=raw_action, role="assistant")
             # parse the raw action
-            action = self.parse_raw_model_response(raw_action)
+            action = self.parse_raw_model_response(raw_action, observation=observation)
 
         # save results and reset
         return {

{plancraft-0.3.6.dist-info → plancraft-0.3.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: plancraft
-Version: 0.3.6
+Version: 0.3.8
 Summary: Plancraft: an evaluation dataset for planning with LLM agents
 License: MIT License
         

{plancraft-0.3.6.dist-info → plancraft-0.3.8.dist-info}/RECORD

@@ -1,6 +1,6 @@
 plancraft/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 plancraft/config.py,sha256=Ppkps-E8xDNYEP9prOVxW2zEG9MpWVzcLJi4tmGLjuQ,4285
-plancraft/evaluator.py,sha256=dTsE3FiQTJc094TmBvfBvefOpGSYcePIGVT36OEIClU,10910
+plancraft/evaluator.py,sha256=FogXAjnldK2K1gD02f2WW1YWvjWokKGUOJxwtwzGNjk,10804
 plancraft/generate_dataset.py,sha256=DlrU-PmvWqSNJD1g1-8Lpb8n3N-Ogw3rje1nrRzjGKs,2382
 plancraft/utils.py,sha256=phaHzbIS85YZrBPaGG9TStHY8ZBKR1LKfuN1exfVy1U,6889
 plancraft/data/test.json,sha256=4jWfYMAVuZCFmGB4iZJAjlh9_8jXECdaGp8xn7_tAM4,1317131
@@ -11,14 +11,14 @@ plancraft/data/val.json,sha256=IToAiaqUNQi_xhX1bzmInuskLaT7C2ryQjP-CZkzL24,13044
 plancraft/data/val.small.easy.json,sha256=9zEmqepjXG2NIp88xnFqOCkwsUsku3HEwHoQGxgTr6U,190252
 plancraft/data/val.small.json,sha256=76E9EFaljDQyAokg97e-IblvcOe6KbrdKkXvRxhhkgo,237653
 plancraft/environment/__init__.py,sha256=XFsFny4lH195AwAmL-WeCaF9ZCMgc7IgXIwhQ8FTdgE,505
-plancraft/environment/actions.py,sha256=D9QqBW7yWsbWCjxNyWp61Xtb0c6EtyXk3PZ1I8SRoBQ,9381
+plancraft/environment/actions.py,sha256=AQxFaK4YW53mPwhuPhHrDF9wENSVjPHSWk0v77I1thw,9460
 plancraft/environment/env.py,sha256=F5xo1eAJ9MeuoE2IpG_LtbaE0BGd66URPB_rehAWIiU,16372
 plancraft/environment/items.py,sha256=Z9rhSyVDEoHF1pxRvhyiT94tyQJaWHi3wUHVcamz82o,221
 plancraft/environment/planner.py,sha256=eJExz3OxSzurIEdH9LOtMwFH9ApqMQ3CokVhmbV6Px0,3953
 plancraft/environment/prompts.py,sha256=8QXclX0ygpL02uZichE1AVkbdn_0HGteD5bzo0FZGOU,6947
 plancraft/environment/recipes.py,sha256=0vwzOU86eZmGN2EpZVSIvzxpx0AOBWNPxTtAOFBN2A0,19570
 plancraft/environment/sampler.py,sha256=IZT-XjmWSZrs0zDyRTMjYytXxewdwYf5YGGdKsR5ll4,7643
-plancraft/environment/search.py,sha256=uFHpLvW40rMKOxDabcyWrpOrhKLDZqAJOF_jew4_WXk,1837
+plancraft/environment/search.py,sha256=Dmdvj04kMvPlwvoWSc2261LTXV8RbMpS4FODV1YoZKs,1847
 plancraft/environment/assets/constants.json,sha256=kyOIOh82CTTMMGEIS60k5k6M-6fkEmYDoGAnvi3Zx5k,1379016
 plancraft/environment/assets/minecraft_font.ttf,sha256=AzoK9cgggXwjFPHtIO7uz-YaDrminl3nvB-VsaTvTAk,60992
 plancraft/environment/assets/table.png,sha256=IKIViZKAPyR4FWnS0JP9AZ19vIEO3qoS5-YRGAO1ow8,5430
@@ -1920,7 +1920,7 @@ plancraft/models/generators.py,sha256=F76_iPiqxUjDIrQwF58tzM0bLM91OkZJ0sBqBuki5w
 plancraft/models/oracle.py,sha256=jDCE6zVFvbwFpDzQZTkHIlRwMud1yMJ4LVIdfpt5ddU,8449
 plancraft/models/utils.py,sha256=E-sZohvolWgGbpHQKgAgkgIfUJoVnT5pMt6JP8xLHKg,4034
 plancraft/train/dataset.py,sha256=oFqEd4LG9oEQ-71teh0Wf7-jJbtybT2ZibfM2bBdBkM,5474
-plancraft-0.3.6.dist-info/METADATA,sha256=MVnklft3zz21a9085j-H-_VYAPJREH2-1ojS08ICbzk,11147
-plancraft-0.3.6.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-plancraft-0.3.6.dist-info/licenses/LICENSE,sha256=YGR8ehDB4t-T-lOQKMfKNR-2zsOU7E3E5NA8t25HKE0,1070
-plancraft-0.3.6.dist-info/RECORD,,
+plancraft-0.3.8.dist-info/METADATA,sha256=H6H7LnIFSd-M-FyqlAtErJfzL4WxSDd0ku66P2YyTeQ,11147
+plancraft-0.3.8.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+plancraft-0.3.8.dist-info/licenses/LICENSE,sha256=YGR8ehDB4t-T-lOQKMfKNR-2zsOU7E3E5NA8t25HKE0,1070
+plancraft-0.3.8.dist-info/RECORD,,