Combined Combinatorial Optimization Dataset

Overview

This dataset is created for evaluating the effectiveness of Large Language Models (LLMs) on various combinatorial optimization problems. Each instance in the dataset represents a problem instance and includes all the necessary attributes to learn and evaluate problem solutions generated by Google's OR Tools.

Problem Types and Attributes

Each instance in the dataset is expected to have the following attributes:

input: The actual problem data (e.g., graph data for VRP, jobs and machines for JSSP, etc.).
instruction: A general description of the problem. This attribute contains problem-specific instructions or details (for example, the number of machines and jobs in a jssp problem). The description varies for each problem type.
output_list_of_list: Expected feasible solution provided as a list of lists.
output_starjob: Expected solution in a human-readable format. This includes detailed explanations such as how the makespan is calculated in JSSP or how the knapsack capacity is updated step by step.

Additionally, instances are tagged with a specific problem_type (e.g., 'vrp', 'jssp', etc.) which may have additional attributes, depending on the specific requirements of the problem type. For example, instances where problem_type equals 'vrp' might include extra attributes relevant only to Vehicle Routing Problems.

Dataset Statistics

Total Instances: 150

Instance Counts by Problem Type

Problem Type	Count
vrp	30
knapsack	30
tsp	30
binpack	30
jssp	30

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:

{'capacity': 'int',
 'city_size': 'int',
 'demands': 'list',
 'input': 'str',
 'instance_id': 'int',
 'instruction': 'str',
 'max_interval': 'int',
 'num_cities': 'int',
 'num_vehicles': 'int',
 'output_list_of_list': 'str',
 'output_starjob': 'str',
 'paired_distances': 'str',
 'problem_type': 'str',
 'time': 'float',
 'vehicle_count': 'int'}

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.