henri24/llm_combinatorial_test

LLM Combinatorial Optimization Test Dataset

This dataset falles under:

Constraint satisfaction problems: Collections teaching models to reason through complex scheduling, resource allocation, or optimization scenarios with multiple interdependent constraints

Overview

This dataset is created for evaluating the effectiveness of Large Language Models (LLMs) in solving various combinatorial optimization problems. The dataset was generated by combining multiple JSON files containing problem instances from different problem domains. Each instance represents a distinct optimization problem along with its associated solution outputs generated using Google's OR Tools.

Structure and Attributes

Each instance in this dataset is a JSON object that contains some common attributes along with problem-specific details. The common attributes present in every instance include:

• input:
  The actual problem data. This attribute represents the raw input needed to describe the optimization problem (e.g., graph information for vehicle routing, job and machine details for job-shop scheduling, etc.).

• instruction:
  A natural language description of the problem. This attribute captures the general instructions and details (such as number of machines and jobs in a job-shop scheduling problem) that vary for each problem type.

• output_list_of_list:
  Expected feasible solution represented as a list of lists. This format outlines a structured solution, potentially describing paths, schedules, or allocations.

• output_starjob:
  A human-readable solution format that includes detailed solution steps. This output explains function calculations and other intermediary steps (for example, how the makespan is computed in a job-shop scheduling problem, or how adjustments are made in knapsack problems).

In addition to these, each instance has the attribute problem_type, which indicates the specific category of the combinatorial optimization problem (such as "vrp", "jssp", etc.). Depending on the problem type, instances may also include extra problem-specific attributes that are unique to that type.

For example:

• Instances with problem_type: "vrp" might contain additional spatial or routing parameters.
• Instances with problem_type: "jssp" include details about the number of machines and jobs, which are also reflected in the instruction attribute.

Attribute Details

Below are the attributes extracted from the first instance along with their inferred data types. Note that the dataset may contain additional problem-specific attributes that vary by problem type:

{'input': 'str',
 'instruction': 'str',
 'makespan': 'int',
 'matrix': 'str',
 'num_jobs': 'int',
 'num_machines': 'int',
 'output_list_of_list': 'str',
 'output_starjob': 'str',
 'path': 'str',
 'problem_type': 'str',
 'prompt_jobs_first': 'str',
 'prompt_machines_first': 'str'}

Conclusion

This dataset provides both structured problem definitions and two types of solution representations to facilitate a comprehensive evaluation of LLM-based approaches across different combinatorial optimization problems.