Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Starjob: Dataset for LLM-Driven Job Shop Scheduling

Henrik Abgaryan, Tristan Cazenave, Ararat Harutyunyan

TL;DR

The Starjob paper investigates applying fine-tuned LLMs to the Job Shop Scheduling Problem (JSSP), an NP-hard domain where makespan minimization is key. It introduces Starjob, a supervised dataset with about 130k natural-language problem/solution pairs, and uses RsLoRA to fine-tune an 8B-parameter LLaMA model on 4-bit quantization, enabling end-to-end scheduling. On Taillard and DMU benchmarks, the fine-tuned Llama model consistently outperforms classic Priority Dispatching Rules and the neural baseline L2D, achieving substantial average gap reductions (e.g., around 21–22% vs 29–38% for baselines). The work demonstrates the viability of LLMs for combinatorial optimization under resource-efficient settings and opens avenues for interactive, NL-based scheduling with potential integration with reinforcement learning and graph-based methods. The Starjob dataset, combined with rsLoRA fine-tuning, provides a practical and scalable pathway to leverage LLMs for complex scheduling tasks in manufacturing and service environments.

Abstract

Large Language Models (LLMs) have shown remarkable capabilities across various domains, but their potential for solving combinatorial optimization problems remains largely unexplored. In this paper, we investigate the applicability of LLMs to the Job Shop Scheduling Problem (JSSP), a classic challenge in combinatorial optimization that requires efficient job allocation to machines to minimize makespan. To this end, we introduce Starjob, the first supervised dataset for JSSP, comprising 130k instances specifically designed for training LLMs. Leveraging this dataset, we fine-tune the LLaMA 8B 4-bit quantized model with the LoRA method to develop an end-to-end scheduling approach. Our evaluation on standard benchmarks demonstrates that the proposed LLM-based method not only surpasses traditional Priority Dispatching Rules (PDRs) but also achieves notable improvements over state-of-the-art neural approaches like L2D, with an average improvement of 15.36% on DMU and 7.85% on Taillard benchmarks. These results highlight the untapped potential of LLMs in tackling combinatorial optimization problems, paving the way for future advancements in this area.

Starjob: Dataset for LLM-Driven Job Shop Scheduling

TL;DR

The Starjob paper investigates applying fine-tuned LLMs to the Job Shop Scheduling Problem (JSSP), an NP-hard domain where makespan minimization is key. It introduces Starjob, a supervised dataset with about 130k natural-language problem/solution pairs, and uses RsLoRA to fine-tune an 8B-parameter LLaMA model on 4-bit quantization, enabling end-to-end scheduling. On Taillard and DMU benchmarks, the fine-tuned Llama model consistently outperforms classic Priority Dispatching Rules and the neural baseline L2D, achieving substantial average gap reductions (e.g., around 21–22% vs 29–38% for baselines). The work demonstrates the viability of LLMs for combinatorial optimization under resource-efficient settings and opens avenues for interactive, NL-based scheduling with potential integration with reinforcement learning and graph-based methods. The Starjob dataset, combined with rsLoRA fine-tuning, provides a practical and scalable pathway to leverage LLMs for complex scheduling tasks in manufacturing and service environments.

Abstract

Large Language Models (LLMs) have shown remarkable capabilities across various domains, but their potential for solving combinatorial optimization problems remains largely unexplored. In this paper, we investigate the applicability of LLMs to the Job Shop Scheduling Problem (JSSP), a classic challenge in combinatorial optimization that requires efficient job allocation to machines to minimize makespan. To this end, we introduce Starjob, the first supervised dataset for JSSP, comprising 130k instances specifically designed for training LLMs. Leveraging this dataset, we fine-tune the LLaMA 8B 4-bit quantized model with the LoRA method to develop an end-to-end scheduling approach. Our evaluation on standard benchmarks demonstrates that the proposed LLM-based method not only surpasses traditional Priority Dispatching Rules (PDRs) but also achieves notable improvements over state-of-the-art neural approaches like L2D, with an average improvement of 15.36% on DMU and 7.85% on Taillard benchmarks. These results highlight the untapped potential of LLMs in tackling combinatorial optimization problems, paving the way for future advancements in this area.

Paper Structure

This paper contains 19 sections, 9 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminary
  4. Dataset Generation
  5. Converting JSSP problem instance to Natural Language: Feature Generation
  6. Definitions
  7. Proposed Method
  8. Description of Rank-Stabilized LoRA Training
  9. Low-Rank Parameter Decomposition
  10. Rank-Stabilization via the Scaling Factor
  11. Loss Function for Fine-Tuning
  12. Zero-shot inference and Label generation
  13. Training Details
  14. Evaluation
  15. Overview of JSSP Solution Parsing and Validation
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Comparison of different methods on TAItaillard1993benchmarks benchmark.
  • Figure 2: Comparison of different methods on DMUdmu_dataset benchmark.
  • Figure 3: Comparison of different methods on TAItaillard1993benchmarks benchmark.
  • Figure 4: Comparison of different methods on DMUdmu_dataset benchmark.