fix some warnings and minor bugs

prover/cli_ruletaker_single_shot_t5-large.yaml

@@ -19,8 +19,8 @@ model:
   model_name: t5-large
   num_beams: 10
   topk: 10
-  verifier_ckpt: ""
-  verifier_weight: 0
+  verifier_ckpt: null
+  verifier_weight: 0.0
   proof_search: false
   oracle_prover: false
   oracle_verifier: false

prover/cli_ruletaker_stepwise_t5-large.yaml

@@ -19,8 +19,8 @@ model:
   model_name: t5-large
   num_beams: 10
   topk: 10
-  verifier_ckpt: ""
-  verifier_weight: 0
+  verifier_ckpt: null
+  verifier_weight: 0.0
   proof_search: false
   oracle_prover: false
   oracle_verifier: false

prover/cli_task1_single_shot_t5-large.yaml

@@ -18,8 +18,8 @@ model:
   model_name: t5-large
   num_beams: 10
   topk: 10
-  verifier_ckpt: ""
-  verifier_weight: 0
+  verifier_ckpt: null
+  verifier_weight: 0.0
   proof_search: false
   oracle_prover: false
   oracle_verifier: false

prover/cli_task1_stepwise_t5-large.yaml

@@ -18,8 +18,8 @@ model:
   model_name: t5-large
   num_beams: 10
   topk: 10
-  verifier_ckpt: ""
-  verifier_weight: 0
+  verifier_ckpt: null
+  verifier_weight: 0.0
   proof_search: false
   oracle_prover: false
   oracle_verifier: false

prover/cli_task2_single_shot_t5-large.yaml

@@ -18,8 +18,8 @@ model:
   model_name: t5-large
   num_beams: 10
   topk: 10
-  verifier_ckpt: ""
-  verifier_weight: 0
+  verifier_ckpt: null
+  verifier_weight: 0.0
   proof_search: false
   oracle_prover: false
   oracle_verifier: false

prover/cli_task2_stepwise_t5-large.yaml

@@ -18,8 +18,8 @@ model:
   model_name: t5-large
   num_beams: 10
   topk: 10
-  verifier_ckpt: ""
-  verifier_weight: 0
+  verifier_ckpt: null
+  verifier_weight: 0.0
   proof_search: false
   oracle_prover: false
   oracle_verifier: false

prover/datamodule.py

@@ -92,9 +92,13 @@ def read_ruletaker_proofs(path: str, is_train: bool) -> List[Example]:
                     "all_proofs": ex["proofs"],
                 }
             )
+            if ex["answer"] == "Unknown":
+                ans = "Unknown"
+            else:
+                ans = not ex["answer"]
             data.append(
                 {
-                    "answer": not ex["answer"] if ex["answer"] != "Unknown" else "Unknown",
+                    "answer": ans,
                     "depth": ex["depth"],
                     "proof": Proof(
                         context,
@@ -133,7 +137,7 @@ def __init__(
         is_train: bool,
     ) -> None:
         super().__init__()
-        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length=512)
         self.max_input_len = max_input_len
         self.max_output_len = max_output_len
         self.is_train = is_train
@@ -207,7 +211,7 @@ def __init__(
         is_train: bool,
     ) -> None:
         super().__init__()
-        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length=512)
         self.max_input_len = max_input_len
         self.max_output_len = max_output_len
         self.sample_goal = sample_goal

prover/model.py

@@ -109,11 +109,11 @@ def __init__(
         warmup_steps: int,
         num_beams: int,
         topk: int,
-        verifier_ckpt: str,
-        verifier_weight: float,
         proof_search: bool,
-        oracle_prover: bool,
-        oracle_verifier: bool,
+        verifier_weight: float,
+        verifier_ckpt: Optional[str] = None,
+        oracle_prover: Optional[bool] = False,
+        oracle_verifier: Optional[bool] = False,
     ) -> None:
         super().__init__()
         self.save_hyperparameters()
@@ -129,12 +129,13 @@ def __init__(
         self.oracle_prover = oracle_prover
         self.oracle_verifier = oracle_verifier
         if stepwise and verifier_weight > 0:
-            assert verifier_weight <= 1
+            assert verifier_weight <= 1.0
+            assert verifier_ckpt is not None
             self.verifiers = [
                 EntailmentClassifier.load_from_checkpoint(verifier_ckpt)
             ]  # Avoid making the verifier a submodule.
 
-        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length=512)
         if (
             model_name.startswith("t5-")
             or model_name.startswith("google/t5-v1_1-")

verifier/cli_entailmentbank_task1.yaml

@@ -10,7 +10,7 @@ model:
   lr: 1e-5
   warmup_steps: 2000
   model_name: roberta-large
-  pos_weight: 128
+  pos_weight: 128.0
 data:
   dataset: entailmentbank
   batch_size: 128

verifier/cli_entailmentbank_task2.yaml

@@ -10,7 +10,7 @@ model:
   lr: 1e-5
   warmup_steps: 2500
   model_name: roberta-large
-  pos_weight: 128
+  pos_weight: 128.0
 data:
   dataset: entailmentbank
   batch_size: 128

verifier/cli_ruletaker.yaml

@@ -10,7 +10,7 @@ model:
   lr: 5e-6
   warmup_steps: 2500
   model_name: roberta-large
-  pos_weight: 4
+  pos_weight: 4.0
 data:
   dataset: ruletaker
   batch_size: 128