World Machine - Data & Report - Toy1D - Experiment 2 Best Long

In the previous experiment, Experiment 1 Configuration Test, we found the best model configuration for a World Machine trained on the Toy1D dataset. However, a question that arises after that is: what is the maximum performance we can achieve with this configuration? In this experiment, we investigate what happens when we train the model for longer. Also, we took this opportunity to explore the effect of the learning rate scheduler used, Cosine Annealing with Warmup. World Machine is a research project that investigates the concept and creation of computational world models. These AI systems create internal representations to understand and make predictions about the external world. See the project page for more information. The project is part of the H.IAAC, the Hub for Artificial Intelligence and Cognitive Architecture, located at the Universidade Estadual de Campinas (UNICAMP), Brazil. The files in this registry are organized by file extension. Each extension contains: json: metrics and logits metadata memmap: logits data pt: trained models png: plots svg: figures with experiment pipelines final_results: specific files of the final results txt: verification files