feat: Add ablation testing

examples_search/miniF2F_search.py

@@ -1,8 +1,9 @@
 #!/usr/bin/env python3
 
-import subprocess, json
+import subprocess, json, argparse
 from pathlib import Path
 from pantograph.server import Server
+from pantograph.search import SearchResult
 from pantograph.search_llm import LLMAgent
 
 def get_project_and_lean_path():
@@ -15,7 +16,7 @@ def read_test_data():
     with open(jsonl_path, 'r') as f:
         return [json.loads(l) for l in list(f)]
 
-def try_test_data(server, agent, entry) -> bool:
+def try_test_data(server, agent, entry: dict, max_steps: int) -> SearchResult:
     e = entry["formal_statement"]
     informal_stmt = entry["informal_stmt"]
     informal_proof = entry["informal_proof"]
@@ -24,14 +25,37 @@ def try_test_data(server, agent, entry) -> bool:
     target = "forall " + ','.join(e.rsplit(':', 1))
     print(f"Target: {target}")
     agent = LLMAgent(server)
-    return agent.search(server=server, target=target, informal_stmt = informal_stmt, informal_proof = informal_proof,verbose=True)
+    return agent.search(server=server, target=target, informal_stmt = informal_stmt, informal_proof = informal_proof,verbose=True, max_steps=max_steps)
+
+def output_file_name(datum, use_hammer: bool, use_llm: bool):
+    name = datum["id"]
+    folder = 'output'
+    if use_hammer:
+        folder += '-hammer'
+    if use_llm:
+        folder += '-llm'
+    folder = Path(__file__).parent / folder
+    folder.mkdir(exist_ok=True, parents=True)
+    return folder / f"{name}.json"
 
 if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+                    prog='MiniF2F Search',
+                    description='Executes LLM on MiniF2F Search')
+    parser.add_argument('--use-hammer', action='store_true')
+    parser.add_argument('--use-llm', action='store_true')
+    parser.add_argument('-s', '--max-steps', default=1000)
+    args = parser.parse_args()
+
     project_path, lean_path = get_project_and_lean_path()
     print(f"$PWD: {project_path}")
     print(f"$LEAN_PATH: {lean_path}")
 
     test_data = read_test_data()
     server = Server(imports=["Mathlib"], project_path=project_path, lean_path=lean_path)
-    agent = LLMAgent(server)
+    agent = LLMAgent(server, use_hammer=args.use_hammer, use_llm=args.use_llm)
-    try_test_data(server, agent, test_data[0])
+    for datum in test_data[:1]:
+        result = try_test_data(server, agent, datum, max_steps=args.max_steps)
+        file_name = output_file_name(datum, args.use_hammer, args.use_llm)
+        with open(file_name, 'w') as f:
+            json.dump({ 'id': datum['id'], 'success': result.success, 'steps': result.steps  }, f)

pantograph/search.py

@@ -26,6 +26,11 @@ class SearchState:
     def is_solved(self) -> bool:
         return all(self.solved)
 
+@dataclass(frozen=True)
+class SearchResult:
+
+    success: bool
+    steps: int
 
 class Agent:
 
@@ -51,7 +56,7 @@ class Agent:
                informal_stmt: str = "",
                informal_proof: str = "",
                max_steps: int = 1000,
-               verbose: bool = False) -> bool:
+               verbose: bool = False) -> SearchResult:
 
         search_stack = [SearchState(state=server.goal_start(target),
                                     parent=None,
@@ -76,7 +81,7 @@ class Agent:
                     if verbose:
                         print("Search complete: Root state solved")
                     self.reset()
-                    return True
+                    return SearchResult(success=True, steps=i_step)
 
                 search_stack.pop(-1)
                 assert not search_stack[search_state.parent].solved[search_state.parent_goal_id]
@@ -97,7 +102,7 @@ class Agent:
                     if verbose:
                         print("Tactic list has been exhausted")
                     self.reset()
-                    return False
+                    return SearchResult(success=False, steps=i_step)
                 continue
 
             try:
@@ -123,7 +128,7 @@ class Agent:
             print("Search iteration limit exhausted")
 
         self.reset()
-        return False
+        return SearchResult(success=False, steps=max_steps)
 
 
 class DumbAgent(Agent):

pantograph/search_llm.py

@@ -8,26 +8,35 @@ import sglang as sgl
 
 class LLMAgent(Agent):
 
-    def __init__(self, server):
+    def __init__(self, server,
+                 use_hammer=True, use_llm=True):
         super().__init__()
         self.n_trials = 5
         self.server = server
         sgl.set_default_backend(sgl.OpenAI("gpt-4"))
 
         self.goal_tactic_id_map = collections.defaultdict(lambda : 0)
+
+        self.use_hammer = use_hammer
+        self.use_llm = use_llm
+        if use_hammer:
             self.tactics = [
                 "aesop",
                 #"simp",
                 #"rfl",
                 #"decide",
             ]
+        else:
+            self.tactics = []
 
     def next_tactic(self, state: GoalState, goal_id: int, informal_stmt:str="",  informal_proof:str="") -> Optional[Tactic]:
         key = (state.state_id, goal_id)
         i = self.goal_tactic_id_map[key]
 
         target = state.goals[goal_id].target
-        if i >= len(self.tactics):
+        if i >= len(self.tactics) and not self.use_llm:
+            return None
+        elif i >= len(self.tactics):
             new_state = None
             for ii in range(self.n_trials):
                 print(f"===============trail {str(ii)}============")