Navigation-trained recurrent organization facilitates relational learning: code, models and data

This dataset accompanies the study “Navigation-trained recurrent organization facilitates relational learning.” It provides trained network weights and underlying experimental results for reproducing the study’s analyses and figures. The deposit contains: P1_pretrained_checkpoints.tar: 20 pretrained recurrent networks—10 navigation-trained networks and 10 delayed-reconstruction controls—with accompanying training records. P2_target_checkpoints.tar: 480 frozen-core target-model checkpoints for the 8 × 8 relational transfer and mechanistic experiments. The frozen cores are reconstructed from P1 and recorded seeds using the accompanying code. P3_source_level_results.tar.gz: Source-, condition-, and target-level experimental results, learning curves, configurations, protocols, and data-splitting records. DATA_LICENSE.md: License terms for the models and data. The accompanying GitHub repository provides executable code, random seeds, data-generation and splitting procedures, statistical and plotting scripts, frozen analysis plans, environment requirements, and configurations. Figure source data and commands for reproducing Figures 2–7 and Supplementary Figures S1–S12 are included. Code and reproduction instructions:https://github.com/NJUTBP/navigation-recurrent-transfer Models and data are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Software in the accompanying GitHub repository is licensed under the MIT License.