Principle:Volcengine Verl Answer Extraction

Knowledge Sources	verl
Domains	Data_Engineering, NLP, Evaluation
Last Updated	2026-02-07 14:00 GMT

Overview

The process of extracting verifiable ground truth answers from raw dataset solution strings using dataset-specific parsing (regex, LaTeX parsing, or direct indexing).

Description

Answer Extraction parses raw solution strings to extract the final answer that can be used for reward computation. Different datasets require different extraction methods:

• GSM8K: Regex extraction of numeric answer after "#### " marker
• MATH: LaTeX \boxed{} parsing to extract the final boxed answer
• HellaSwag: Direct integer label indexing for multiple-choice
• Geo3K: Choice letter extraction for geometry problems

The extracted answer becomes the ground_truth field in the reward configuration, used during training to compare against model-generated answers.

Usage

Use answer extraction during data preprocessing whenever the dataset has verifiable answers that need to be parsed from raw solution text.

Theoretical Basis

Answer extraction is a dataset-specific parsing function:

# Abstract answer extraction
def extract_answer(solution_str, dataset_type):
    if dataset_type == "gsm8k":
        match = re.search(r"#### (\-?[0-9\.\,]+)", solution_str)
        return match.group(1) if match else None
    elif dataset_type == "math":
        return remove_boxed(last_boxed_only_string(solution_str))
    elif dataset_type == "multiple_choice":
        return int(solution_str)  # Index of correct choice

Related Pages

Implemented By

• Implementation:Volcengine_Verl_Extract_Solution_Regex