Implementation:Iamhankai Forest of Thought Extract Label

Knowledge Sources	Forest-of-Thought
Domains	NLP, Evaluation
Last Updated	2026-02-14 03:00 GMT

Overview

Concrete tool for extracting answer labels from LLM output text provided by the Forest-of-Thought repository.

Description

The extract_label function extracts a clean answer label from raw LLM output. For non-GSM datasets, it looks for \\boxed{} content or falls back to plain text extraction. For GSM8K, it extracts the number after #### or The answer is. The function uses regex patterns to find and clean the answer, handling LaTeX formatting and multi-format outputs.

Usage

Called extensively throughout the codebase: in monte_carlo_tree() for per-iteration answer tracking, in process_answer() for CGDM post-processing, and in check() for answer comparison preprocessing.

Code Reference

Source Location

• Repository: Forest-of-Thought
• File: utils/utils.py
• Lines: L199-225

Signature

def extract_label(DATA_NAME, text: str, type='') -> str:
    """
    Extract answer label from model output text.

    Args:
        DATA_NAME (str): Dataset identifier (gsm8k, math, aime).
        text (str): Raw model output containing the answer.
        type (str): Answer type hint ('digit', 'option', 'yesorno',
                    'formula'). Default: ''.

    Returns:
        str: Extracted answer label, or None if not found.
    """

Import

from utils.utils import extract_label

I/O Contract

Inputs

Name	Type	Required	Description
DATA_NAME	str	Yes	Dataset name for format-aware extraction
text	str	Yes	Raw LLM output text
type	str	No	Answer type hint for targeted extraction

Outputs

Name	Type	Description
label	str or None	Extracted answer label, cleaned of formatting

Usage Examples

from utils.utils import extract_label

# MATH format with boxed answer
label = extract_label("math", "The solution is \\boxed{42}.")
# label = "42"

# GSM8K format with ####
label = extract_label("gsm8k", "Step 1: ... Step 2: ... #### 120")
# label = "120"

# Fallback to last number
label = extract_label("math", "The answer is 7 because...")
# label = "7"

Related Pages

Implements Principle

• Principle:Iamhankai_Forest_of_Thought_Answer_Extraction