plancraft 0.3.18__py3-none-any.whl → 0.3.20__py3-none-any.whl

plancraft/evaluator.py

@@ -4,7 +4,6 @@ from typing import Optional
 from copy import deepcopy
 
 import imageio
-from loguru import logger
 from tqdm import tqdm
 
 import wandb
@@ -293,6 +292,8 @@ class Evaluator:
             # Get observations for all active environments
             observations = []
             active_indices = []
+            active_histories = []
+
             for i, (env, action, active) in enumerate(
                 zip(environments, actions, active_mask)
             ):
@@ -311,17 +312,14 @@ class Evaluator:
                         "example_id": examples[i].id,
                         "images": histories[i].images,
                     }
-                    logger.info("STOP")
                     continue
 
-                active_indices.append(i)
+                # Get observation
                 if isinstance(action, str):
-                    # Handle message action
                     obs = env.step()
                     obs["target"] = examples[i].target
                     obs["message"] = action
                 else:
-                    # Handle environment action
                     obs = env.step(action)
                     obs["target"] = examples[i].target
                     obs["message"] = self.convert_observation_to_message(
@@ -342,7 +340,12 @@ class Evaluator:
                         }
                         continue
 
+                # Add to batch lists
+                active_indices.append(i)
                 observations.append(obs)
+                active_histories.append(histories[i])
+
+                # Update history
                 histories[i].add_observation_to_history(obs)
                 histories[i].add_message_to_history(content=obs["message"], role="user")
                 steps_taken[i] += 1
@@ -351,20 +354,20 @@ class Evaluator:
                 break
 
             # Batch predict actions for active environments
-            active_histories = [histories[i] for i in active_indices]
             raw_actions = model.batch_step(
                 observations, dialogue_histories=active_histories
             )
 
             # Process actions for each active environment
-            for idx, raw_action in zip(active_indices, raw_actions):
-                logger.info(f"{histories[idx].num_steps}, {raw_action}")
+            for batch_idx, (idx, raw_action) in enumerate(
+                zip(active_indices, raw_actions)
+            ):
                 histories[idx].add_message_to_history(
                     content=raw_action, role="assistant"
                 )
                 actions[idx] = self.parse_raw_model_response(
                     raw_action,
-                    observation=observations[active_indices.index(idx)],
+                    observation=observations[batch_idx],
                     history=histories[idx],
                 )
 

{plancraft-0.3.18.dist-info → plancraft-0.3.20.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: plancraft
-Version: 0.3.18
+Version: 0.3.20
 Summary: Plancraft: an evaluation dataset for planning with LLM agents
 License: MIT License
         

{plancraft-0.3.18.dist-info → plancraft-0.3.20.dist-info}/RECORD

@@ -1,6 +1,6 @@
 plancraft/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 plancraft/config.py,sha256=ShsFRlJ7plsl3ToD9fiO_4LDQuXdbjNV6Xp6o3Yk2Yg,4315
-plancraft/evaluator.py,sha256=v8itX8buduqTZdR39gtLwdhKGEnSX3rJv9Yd13EzNgQ,16395
+plancraft/evaluator.py,sha256=R_RZN9AL_ae0rIvj7HLhYolTpCVMuhPTJfIrmyoLaX4,16326
 plancraft/generate_dataset.py,sha256=DlrU-PmvWqSNJD1g1-8Lpb8n3N-Ogw3rje1nrRzjGKs,2382
 plancraft/utils.py,sha256=67UUDMSv8TqX_I0fL5-yG_vkHvTZlnhSLkktWAg5p34,5712
 plancraft/data/test.json,sha256=4jWfYMAVuZCFmGB4iZJAjlh9_8jXECdaGp8xn7_tAM4,1317131
@@ -1920,7 +1920,7 @@ plancraft/models/generators.py,sha256=F76_iPiqxUjDIrQwF58tzM0bLM91OkZJ0sBqBuki5w
 plancraft/models/oracle.py,sha256=f-0KWlBuHy6wcxmDsxM3MQ_QwfBstzfbA26mlk1MgLA,1657
 plancraft/models/utils.py,sha256=E-sZohvolWgGbpHQKgAgkgIfUJoVnT5pMt6JP8xLHKg,4034
 plancraft/train/dataset.py,sha256=oFqEd4LG9oEQ-71teh0Wf7-jJbtybT2ZibfM2bBdBkM,5474
-plancraft-0.3.18.dist-info/METADATA,sha256=p_Ln_3jx77ygBZG6yjuLhVs883PysUXUCi1sK67QvJs,11148
-plancraft-0.3.18.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-plancraft-0.3.18.dist-info/licenses/LICENSE,sha256=YGR8ehDB4t-T-lOQKMfKNR-2zsOU7E3E5NA8t25HKE0,1070
-plancraft-0.3.18.dist-info/RECORD,,
+plancraft-0.3.20.dist-info/METADATA,sha256=7Gs3Txfw2YCBVDAstW7K69fho4vktcYp6sw6g2nQOYE,11148
+plancraft-0.3.20.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+plancraft-0.3.20.dist-info/licenses/LICENSE,sha256=YGR8ehDB4t-T-lOQKMfKNR-2zsOU7E3E5NA8t25HKE0,1070
+plancraft-0.3.20.dist-info/RECORD,,