plancraft 0.3.3__py3-none-any.whl → 0.3.4__py3-none-any.whl

plancraft/environment/prompts.py

@@ -59,7 +59,8 @@ SEARCH_STEPS = [
 
 def get_system_prompt(
     handlers: list[ActionHandlerBase] = [MoveActionHandler(), SmeltActionHandler()],
-):
+    use_multimodal_content_format=False,
+) -> dict:
     action_names = [handler.action_name for handler in handlers]
     assert "move" in action_names, "MoveActionHandler should be one of the handlers"
     assert "smelt" in action_names, "SmeltActionHandler should be one of the handlers"
@@ -72,7 +73,17 @@ def get_system_prompt(
     for handler in handlers:
         output_format += f"\n\t- {handler.prompt_format_example}"
 
-    return f"{BASE_SYSTEM_PROMPT}\n\nActions:{descriptions}\n\nFormat{output_format}\n\n{BASE_SYSTEM_PROMPT_EXAMPLE}"
+    system_prompt_text = f"{BASE_SYSTEM_PROMPT}\n\nActions:{descriptions}\n\nFormat{output_format}\n\n{BASE_SYSTEM_PROMPT_EXAMPLE}"
+
+    if use_multimodal_content_format:
+        return {
+            "role": "system",
+            "content": [{"text": system_prompt_text, "type": "text"}],
+        }
+    return {
+        "role": "system",
+        "content": system_prompt_text,
+    }
 
 
 def get_prompt_example(

plancraft/evaluator.py

@@ -1,5 +1,6 @@
 import json
 import os
+from typing import Optional
 
 import imageio
 from loguru import logger
@@ -8,18 +9,18 @@ from tqdm import tqdm
 import wandb
 from plancraft.config import PlancraftExample
 from plancraft.environment.actions import (
-    StopAction,
     ActionHandlerBase,
     MoveActionHandler,
     SmeltActionHandler,
+    StopAction,
 )
 from plancraft.environment.env import (
     PlancraftEnvironment,
     get_objective_str,
     target_and_inventory_to_text_obs,
 )
-from plancraft.utils import History
 from plancraft.models.base import PlancraftBaseModel
+from plancraft.utils import History
 
 
 class Evaluator:
@@ -48,6 +49,10 @@ class Evaluator:
         use_images: bool = False,
         use_text_inventory: bool = False,
         use_fasterrcnn: bool = False,
+        system_prompt: Optional[dict] = None,
+        prompt_examples: list[dict] = [],
+        prompt_images: list[str] = [],
+        few_shot: bool = True,
     ):
         self.run_name = run_name
         self.use_multimodal_content_format = use_multimodal_content_format
@@ -77,6 +82,10 @@ class Evaluator:
             use_images=use_images,
             use_text_inventory=use_text_inventory,
             resolution=resolution,
+            few_shot=few_shot,
+            system_prompt=system_prompt,
+            prompt_examples=prompt_examples,
+            prompt_images=prompt_images,
         )
 
         # load model

plancraft/utils.py

@@ -2,6 +2,7 @@ import glob
 import pathlib
 from collections import Counter
 from copy import copy
+from typing import Optional
 
 import torch
 from loguru import logger
@@ -12,8 +13,8 @@ from plancraft.environment.actions import (
     SmeltAction,
 )
 from plancraft.environment.prompts import (
-    get_system_prompt,
     get_prompt_example,
+    get_system_prompt,
     load_prompt_images,
 )
 
@@ -35,6 +36,9 @@ class History:
         use_images=False,
         use_text_inventory=False,
         resolution="high",
+        system_prompt: Optional[dict] = None,
+        prompt_examples: list[dict] = [],
+        prompt_images: list[str] = [],
     ):
         self.action_handlers = actions
         self.use_multimodal_content_format = use_multimodal_content_format
@@ -49,31 +53,30 @@ class History:
 
         self.tokens_used = 0
 
+        # use system prompt if provided
+        if system_prompt:
+            self.system_prompt_dialogue = system_prompt
+        else:
+            # generate system prompt
+            self.system_prompt_dialogue = get_system_prompt(
+                handlers=self.action_handlers,
+                use_multimodal_content_format=self.use_multimodal_content_format,
+            )
+
         # set up dialogue history with few-shot prompt
+        self.prompt_examples = prompt_examples
+        self.prompt_images = prompt_images
         self.set_up_few_shot_prompt()
-        self.system_prompt_dialogue = self.system_prompt()
 
         self.dialogue_history = copy(self.prompt_examples)
         self.images = copy(self.prompt_images)
         self.initial_dialogue_length = len(self.dialogue_history)
 
-    def system_prompt(self):
-        # kept separate from dialogue history because certain models deal with system prompt differently
-        system_prompt_text = get_system_prompt(handlers=self.action_handlers)
-        if self.use_multimodal_content_format:
-            return {
-                "role": "system",
-                "content": [{"text": system_prompt_text, "type": "text"}],
-            }
-        return {
-            "role": "system",
-            "content": system_prompt_text,
-        }
-
     def set_up_few_shot_prompt(self):
-        self.prompt_examples = []
-        self.prompt_images = []
-
+        # if either prompt_examples or prompt_images are provided, skip
+        if self.prompt_examples or self.prompt_images:
+            return
+        # if few-shot is not enabled, skip
         if self.few_shot:
             self.prompt_examples = get_prompt_example(
                 self.action_handlers,

{plancraft-0.3.3.dist-info → plancraft-0.3.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: plancraft
-Version: 0.3.3
+Version: 0.3.4
 Summary: Plancraft: an evaluation dataset for planning with LLM agents
 License: MIT License
         

{plancraft-0.3.3.dist-info → plancraft-0.3.4.dist-info}/RECORD

@@ -1,8 +1,8 @@
 plancraft/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 plancraft/config.py,sha256=Ppkps-E8xDNYEP9prOVxW2zEG9MpWVzcLJi4tmGLjuQ,4285
-plancraft/evaluator.py,sha256=7PjdITOTUCtjPywFOOd9vVhl5UDKZuFF7rc3mjUOn0Q,10717
+plancraft/evaluator.py,sha256=adGmrn3GMQd5KSfFGQZxHjisQbvoxvEv1W1CPxZnFi8,11061
 plancraft/generate_dataset.py,sha256=DlrU-PmvWqSNJD1g1-8Lpb8n3N-Ogw3rje1nrRzjGKs,2382
-plancraft/utils.py,sha256=7VWKVlDhoMacRypRRSKM1K3hwwJ0nHR3zyx9jZH1C1g,7042
+plancraft/utils.py,sha256=rYiqLUaEqjdUG-nqeHmeVG3PaExAlYiBGXH5qzLZPhs,7224
 plancraft/data/test.json,sha256=4jWfYMAVuZCFmGB4iZJAjlh9_8jXECdaGp8xn7_tAM4,1317131
 plancraft/data/test.small.easy.json,sha256=5NZEJ2PqIgmHQecJOIVQyM1D6GFKyJq7GVmgRudaqQk,189304
 plancraft/data/test.small.json,sha256=eULAG1rdolRMXPrecV-7YoDIheKGyIT5MVpWdISV0wg,270089
@@ -15,7 +15,7 @@ plancraft/environment/actions.py,sha256=D9QqBW7yWsbWCjxNyWp61Xtb0c6EtyXk3PZ1I8SR
 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=OKxiv02NIhRk5FZJUEDRLkVWVMc-aXKJi7i7X61uUmk,6633
+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
@@ -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-0.3.3.dist-info/METADATA,sha256=UMIYGLhvaJO8CJyOj4hka_5OO2T728yqhzhX0rMQqfQ,11306
-plancraft-0.3.3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-plancraft-0.3.3.dist-info/licenses/LICENSE,sha256=YGR8ehDB4t-T-lOQKMfKNR-2zsOU7E3E5NA8t25HKE0,1070
-plancraft-0.3.3.dist-info/RECORD,,
+plancraft-0.3.4.dist-info/METADATA,sha256=W14g4fJ1y6zALGre8NKFRZXu9cVCrQS9i-24akOIWSw,11306
+plancraft-0.3.4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+plancraft-0.3.4.dist-info/licenses/LICENSE,sha256=YGR8ehDB4t-T-lOQKMfKNR-2zsOU7E3E5NA8t25HKE0,1070
+plancraft-0.3.4.dist-info/RECORD,,