fix regex

lm_eval/tasks/mathvista/mathvista.yaml

@@ -5,7 +5,7 @@ output_type: "generate_until"
 #process_docs: !function utils.process_docs
 doc_to_image:
   - decoded_image
-doc_to_text: "<image> {{query}}"
+doc_to_text: "<image>{{query}}"
 #doc_to_choice: '{{ ["A", "B", "C", "D", "E", "F"][:choices.length] }}'
 doc_to_target: answer
 process_results: !function utils.process_results
@@ -14,7 +14,7 @@ generation_kwargs:
     - "<|endoftext|>"
   temperature: 0.0
   do_sample: false
-  max_gen_toks: 64
+  max_gen_toks: 1024
 metric_list:
   - metric: acc
     aggregation: mean

lm_eval/tasks/mathvista/utils.py

@@ -66,7 +66,7 @@ def normalize_extracted_answer(
 
     elif answer_type == "float":
         try:
-            normalized_extraction = str(round(float(extraction), precision))
+            normalized_extraction = str(round(float(extraction), int(precision)))
         except Exception:
             normalized_extraction = None
 
@@ -102,7 +102,9 @@ def extract_answer(response: str, problem: dict) -> str:
         return ""
 
     ### This is not in the original code:
-    extract = re.findall(r"[tT]he answer is (\d+)", response)
+    extract = re.findall(
+        r"[tT]he answer is ([A-Za-z0-9]+(?:\.[A-Za-z0-9]+)?)", response
+    )
     if extract:
         return str(extract[0])
     ###
@@ -124,7 +126,7 @@ def extract_answer(response: str, problem: dict) -> str:
         except Exception:
             pass
 
-    return ""
+    return response
 
 
 # adapted from https://github.com/lupantech/MathVista/blob/main/evaluation/extract_answer.py