metadata

license: cc
task_categories:
  - question-answering
language:
  - en
tags:
  - code
  - math
  - reasoning
  - combinatorial_problems
  - optimization
pretty_name: SearchBench
size_categories:
  - 1K<n<10K

Dataset Card for SearchBench

Dataset Description
- Dataset Summary
- Languages
Dataset Structure
- Data Instances
- Data Fields
Additional Information
- Licensing Information
- Citation Information

Dataset Description

Repository: https://github.com/NasimBorazjani/Navigating_Labyrinth.git
Paper:
Point of Contact: nasimborazjani@berkeley.edu & nb11@williams.edu

Dataset Summary

SearchBench is a benchmark designed to evaluate Language Models' (LLMs) ability to solve state-based problems that require combinatorial search and backtracking. SearchBench problems require a systematic exploration of action paths and backtracking to feasible states, which poses a significant challenge for LLMs to solve end-to-end, due to their autoregressive next-token prediction architecture.

The dataset is composed of five problem categories: puzzles, subset sum, sorting, pathfinding, and under-determined systems. These categories are further divided into 11 distinct problem types. Each problem type is adapted from NP-hard puzzles and combinatorial problems, with modifications made to the rules and constraints of each problem. The dataset contains approximately 100 instances of each problem type with varying levels of difficulty.

Automated pipelines are implemented for each problem type to generate an unlimited number of solvable instances and to assess the feasibility, correctness, and optimality of the solutions generated by the LLMs.

This benchmark provides insights into LLMs' capacity to implement new algorithms to solve complex problems. It also investigates the non-linear reasoning capability of LLMs to solve search problems end-to-end using only text.

Languages

The text in the dataset is in English. The associated BCP-47 code is en.

Dataset Structure

Data Instances

Each instance contains a natural language description of a problem and the optimal solution for the problem calculated using an A* algorithm with provable admissible and consistent heuristic implementation.

{
    “problem_statement”: “In the game 'Sort the Chars', we are given a table of n by m dimensions. This table contains n words, each with m characters, except for the first word which has m - 1 characters. Each character is written on a separate tile. The objective of the game is to rearrange the characters such that row i spells the i-th word in the list, with the blank tile ('_') placed in the top left corner of the board in the end. We can rearrange the tiles by swapping the blank space with any of its 4 diagonal neighboring tiles. Given the list of words and initial state of the board below, where the black space is represented as '_', what is the shortest list of swap actions (reported in python syntax) that can sort the board into the given list of target words? The list must only include the 4 diagonal swap directions: up-right, down-right, up-left, or down-left, representing the direction in ehich the blank space was swpped in. Target words: cam, hill, pray, doer   The initial board: [['i', 'c', 'a', 'm'], ['h', 'p', 'l', 'o'], ['_', 'r', 'a', 'y'], ['d', 'l', 'e', 'r']]”,
    “opt_solution”: "['up-right', 'down-right', 'down-left', 'up-left', 'up-right', 'down-right', 'up-right', 'up-left', 'down-left', 'down-left', 'down-right', 'up-right', 'up-right', 'up-left', 'down-left', 'up-left']",
}

Data Fields

Each instance in the dataset comes with several additional fields. These fields offer insights about the problem and allow for automatic evaluation of the solutions generated by the LLMs across multiple dimensions. For a more detailed description of the fields for each instance, please refer to the supplementary material of the paper.

diff_sorted_id: A unique identifier for each problem, sorted by difficulty level within a specific problem type.
problem_statement: The description of the problem to be solved, given to language models.
problem_type: The type of problem, out of 11 possible types.
problem_category: The category of the problem, out of five possible categories.
relative_diff_score: A score indicating the problem's difficulty relative to other problems of the same type.
opt_solution: A list of actions leading to the goal state with the minimum cost.
opt_solution_cost: The cost of the optimal solution for the problem.
opt_solution_compute_t: The time taken by the A* implementation to solve the problem.
solution_depth: The number of actions required to reach the goal state with minimum cost.
max_successor_states: The maximum number of successor states that can be reached from any state in the problem.
num_vars_per_state: An upper limit on the number of variables in each state of the problem.
is_feasible_args: Variables that must be passed to the 'is_feasible' function to check if a solution adheres to the problem's rules and constraints.
is_correct_args: Variables that must be passed to the 'is_correct' function to evaluate the correctness of a solution.
A*_args: Variables that must be passed to the A* implementation to obtain the optimal solution for the problem.

Additional Information

Licensing Information

The SearchBench dataset is licensed under the CC BY-SA.

Datasets:

NasimBrz
/

SearchBench