Upgraded to dspy-ai==2.5.7 resolved ValueError via inclusion of basic prompt instruction for critique signature and typed CoT

Author: NumberChiffre

mcts_llm/mctsr.py

@@ -45,6 +45,14 @@ class CritiqueAnswer(dspy.Signature):
 
 
 class RefineAnswer(dspy.Signature):
+    """[[ ## proposed_instruction ## ]] Given a mathematical problem, a current answer, and a critique of that answer,
+    refine the current answer to provide a more accurate and well-reasoned solution. Begin by carefully analyzing the
+    problem and the critique, then think step by step to derive the correct answer. Ensure that your reasoning is clear
+    and logical, and that the final answer is justified by the steps taken.
+
+    [[ ## completed ## ]]
+    """
+
     problem: str = dspy.InputField()
     current_answer: str = dspy.InputField()
     critique: str = dspy.InputField()
@@ -59,7 +67,7 @@ class EvaluateAnswer(dspy.Signature):
 
 class ZeroShotCoT(dspy.Module):
     def __init__(self):
-        self.cot = dspy.ChainOfThought(ZeroShotAnswer)
+        self.cot = dspy.TypedChainOfThought(ZeroShotAnswer)
 
     def forward(self, problem) -> dspy.Prediction:
         return dspy.Prediction(answer=self.cot(problem=problem).answer)
@@ -69,8 +77,8 @@ class MultipleTurnSelfRefine(dspy.Module):
     def __init__(self, num_turns: int = 1):
         super().__init__()
         self.zero_shot_cot = ZeroShotCoT()
-        self.critique_answer = dspy.ChainOfThought(CritiqueAnswer)
-        self.refine_answer = dspy.ChainOfThought(RefineAnswer)
+        self.critique_answer = dspy.TypedChainOfThought(CritiqueAnswer)
+        self.refine_answer = dspy.TypedChainOfThought(RefineAnswer)
         self.num_turns = num_turns
 
     def forward(self, problem) -> dspy.Prediction:
@@ -140,9 +148,9 @@ def __init__(
         self.samples_per_node = samples_per_node
 
         self.zero_shot = ZeroShotCoT()
-        self.critique = dspy.ChainOfThought(CritiqueAnswer)
-        self.evaluate = dspy.ChainOfThought(EvaluateAnswer)
-        self.refine = dspy.ChainOfThought(RefineAnswer)
+        self.critique = dspy.TypedChainOfThought(CritiqueAnswer)
+        self.evaluate = dspy.TypedChainOfThought(EvaluateAnswer)
+        self.refine = dspy.TypedChainOfThought(RefineAnswer)
 
     def initialize(self, S: MCTSrState) -> MCTSrNode:
         if self.initialize_strategy == InitializeStrategy.ZERO_SHOT:

notebooks/evaluate_qwen25-7b-instruct.ipynb

@@ -17,7 +17,7 @@ classifiers = [
 
 [tool.poetry.dependencies]
 python = ">=3.12,<3.13"
-dspy-ai = "2.4.17"
+dspy-ai = "^2.5.6"
 python-dotenv = "^1.0.1"
 
 [tool.poetry.group.dev.dependencies]

notebooks/finetune_qwen25-7b-instruct.ipynb

@@ -24,7 +24,7 @@ def mock_parse_integer_answer():
 
 @pytest.fixture
 def mock_chain_of_thought():
-    with patch("dspy.ChainOfThought") as mock:
+    with patch("dspy.TypedChainOfThought") as mock:
         mock_instance = Mock()
         mock_instance.return_value = Mock(answer="Mocked answer", critique="Mock critique")
         mock.return_value = mock_instance
@@ -57,7 +57,7 @@ def mctsr(mock_parse_integer_answer, mock_chain_of_thought):
 
 
 def test_zero_shot_cot():
-    with patch("mcts_llm.mctsr.dspy.ChainOfThought") as mock_chain_of_thought:
+    with patch("mcts_llm.mctsr.dspy.TypedChainOfThought") as mock_chain_of_thought:
         mock_cot = Mock()
         mock_chain_of_thought.return_value = mock_cot
         mock_cot.return_value = dspy.Prediction(answer="Test answer")
@@ -72,7 +72,7 @@ def test_zero_shot_cot():
 
 @pytest.mark.parametrize("num_turns", [1, 3])
 def test_multiple_turn_self_refine(num_turns):
-    with patch("mcts_llm.mctsr.dspy.ChainOfThought") as mock_chain_of_thought:
+    with patch("mcts_llm.mctsr.dspy.TypedChainOfThought") as mock_chain_of_thought:
         mock_zero_shot = Mock()
         mock_critique = Mock()
         mock_refine = Mock()