suhaas-code commited on
Commit
76acce0
·
1 Parent(s): ed507c1

phase2-graders-2

Browse files
Files changed (2) hide show
  1. openenv.yaml +3 -0
  2. tasks/task_definitions.py +8 -2
openenv.yaml CHANGED
@@ -72,9 +72,12 @@ tasks:
72
  - id: task_easy_yield
73
  difficulty: easy
74
  description: Maximize agronomic yield performance using normalized total reward.
 
75
  - id: task_medium_chemical_efficiency
76
  difficulty: medium
77
  description: Minimize total fertilizer and pesticide while maintaining productivity.
 
78
  - id: task_hard_sustainability_balance
79
  difficulty: hard
80
  description: Optimize the long-term yield-to-chemical-use balance ratio.
 
 
72
  - id: task_easy_yield
73
  difficulty: easy
74
  description: Maximize agronomic yield performance using normalized total reward.
75
+ grader: grade_yield_performance
76
  - id: task_medium_chemical_efficiency
77
  difficulty: medium
78
  description: Minimize total fertilizer and pesticide while maintaining productivity.
79
+ grader: grade_chemical_efficiency
80
  - id: task_hard_sustainability_balance
81
  difficulty: hard
82
  description: Optimize the long-term yield-to-chemical-use balance ratio.
83
+ grader: grade_sustainability_balance
tasks/task_definitions.py CHANGED
@@ -10,19 +10,21 @@ from typing import Dict, List, Any
10
 
11
  class TaskDefinition:
12
  """Represents a task with metadata and grader association."""
13
-
14
  def __init__(
15
  self,
16
  task_id: str,
17
  name: str,
18
  description: str,
19
  difficulty: str,
 
20
  ):
21
  self.task_id = task_id
22
  self.name = name
23
  self.description = description
24
  self.difficulty = difficulty
25
-
 
26
  def to_dict(self) -> Dict[str, str]:
27
  """Convert to OpenEnv task definition format."""
28
  return {
@@ -30,6 +32,7 @@ class TaskDefinition:
30
  "name": self.name,
31
  "description": self.description,
32
  "difficulty": self.difficulty,
 
33
  }
34
 
35
 
@@ -41,6 +44,7 @@ TASKS: List[TaskDefinition] = [
41
  description="Maximize crop yield through optimal irrigation and environmental management. "
42
  "Graded on average reward per step.",
43
  difficulty="easy",
 
44
  ),
45
  TaskDefinition(
46
  task_id="task_medium_chemical_efficiency",
@@ -48,6 +52,7 @@ TASKS: List[TaskDefinition] = [
48
  description="Minimize fertilizer and pesticide usage while maintaining acceptable yields. "
49
  "Graded on chemical use efficiency.",
50
  difficulty="medium",
 
51
  ),
52
  TaskDefinition(
53
  task_id="task_hard_sustainability_balance",
@@ -55,6 +60,7 @@ TASKS: List[TaskDefinition] = [
55
  description="Achieve top-tier sustainability by maximizing yield-to-chemical-input ratio. "
56
  "Graded on sustainability metrics.",
57
  difficulty="hard",
 
58
  ),
59
  ]
60
 
 
10
 
11
  class TaskDefinition:
12
  """Represents a task with metadata and grader association."""
13
+
14
  def __init__(
15
  self,
16
  task_id: str,
17
  name: str,
18
  description: str,
19
  difficulty: str,
20
+ grader: str,
21
  ):
22
  self.task_id = task_id
23
  self.name = name
24
  self.description = description
25
  self.difficulty = difficulty
26
+ self.grader = grader
27
+
28
  def to_dict(self) -> Dict[str, str]:
29
  """Convert to OpenEnv task definition format."""
30
  return {
 
32
  "name": self.name,
33
  "description": self.description,
34
  "difficulty": self.difficulty,
35
+ "grader": self.grader,
36
  }
37
 
38
 
 
44
  description="Maximize crop yield through optimal irrigation and environmental management. "
45
  "Graded on average reward per step.",
46
  difficulty="easy",
47
+ grader="grade_yield_performance",
48
  ),
49
  TaskDefinition(
50
  task_id="task_medium_chemical_efficiency",
 
52
  description="Minimize fertilizer and pesticide usage while maintaining acceptable yields. "
53
  "Graded on chemical use efficiency.",
54
  difficulty="medium",
55
+ grader="grade_chemical_efficiency",
56
  ),
57
  TaskDefinition(
58
  task_id="task_hard_sustainability_balance",
 
60
  description="Achieve top-tier sustainability by maximizing yield-to-chemical-input ratio. "
61
  "Graded on sustainability metrics.",
62
  difficulty="hard",
63
+ grader="grade_sustainability_balance",
64
  ),
65
  ]
66