This paper proposes a self-supervised training strategy for combinatorial problems, specifically the Job Shop Scheduling (JSP) problem. The method, called Self-Labeling Improvement Method (SLIM), trains generative models by sampling multiple solutions and using the best one according to the problem objective as a pseudo-label. This approach eliminates the need for optimality information and relies only on self-supervision. The method is validated on the JSP, a complex combinatorial problem, using a generative model based on the Pointer Network. Experiments on popular benchmarks show that the resulting models outperform constructive heuristics and state-of-the-art learning proposals for the JSP. Additionally, the robustness of SLIM is demonstrated by applying it to the Traveling Salesman Problem (TSP). The key contributions of this work are the introduction of SLIM, a novel self-labeling improvement method for training generative models, and a generative encoder-decoder architecture capable of generating high-quality solutions for JSP instances in seconds. The paper also discusses the methodology, results, and comparisons with other approaches, showing that SLIM is effective in solving the JSP and other combinatorial problems.This paper proposes a self-supervised training strategy for combinatorial problems, specifically the Job Shop Scheduling (JSP) problem. The method, called Self-Labeling Improvement Method (SLIM), trains generative models by sampling multiple solutions and using the best one according to the problem objective as a pseudo-label. This approach eliminates the need for optimality information and relies only on self-supervision. The method is validated on the JSP, a complex combinatorial problem, using a generative model based on the Pointer Network. Experiments on popular benchmarks show that the resulting models outperform constructive heuristics and state-of-the-art learning proposals for the JSP. Additionally, the robustness of SLIM is demonstrated by applying it to the Traveling Salesman Problem (TSP). The key contributions of this work are the introduction of SLIM, a novel self-labeling improvement method for training generative models, and a generative encoder-decoder architecture capable of generating high-quality solutions for JSP instances in seconds. The paper also discusses the methodology, results, and comparisons with other approaches, showing that SLIM is effective in solving the JSP and other combinatorial problems.