A regular expression (regex) searches response text for structured patterns. In answer extraction, it recognizes numeric candidates such as fractions, decimals, and scientific notation when a final boxed answer is unavailable. section 3.4
Use regex-based extraction alongside boxed-answer parsing:
- Compile a numeric pattern with Python’s re library. section 3.4
- Search the full response and select the last matching number. section 3.4
- Apply fallback policies such as number_then_full, number_only, or none. section 3.4
Extracting a final numeric candidate
Given response text containing “abc < > 14/3 abc” and no boxed answer, the regex finds 14/3 as the final candidate. A boxed value such as \boxed{14/3} takes precedence. section 3.4
Listing 3.5 Extracting the final answer candidate
import re
RE_NUMBER = re.compile( #1
r"-?(?:\d+/\d+|\d+(?:\.\d+)?(?:[eE][+-]?\d+)?)"
)
def extract_final_candidate(text, fallback="number_then_full"):
result = "" #2
if text: #3
boxed = get_last_boxed(text.strip())
if boxed:
result = boxed.strip().strip("$ ")
elif fallback in ("number_then_full", "number_only"):#4
m = RE_NUMBER.findall(text)
if m:
result = m[-1] #5
elif fallback == "number_then_full":
result = text #6
return result
Inspect the compiled numeric pattern and fallback implementation used to extract the final candidate. section 3.4
