LEAP: Training Logs and Checkpoints for the 95-Run Experiment Matrix (ChinaXiv 202608.00173)

This dataset contains the complete experimental records for the paper"LEAP: Lagrangian Dual-Ascent Safety Constraints for Latent-Space ModelPredictive Control" (ChinaXiv: 202608.00173,chinaxiv.org/abs/202608.00173).Archive: leap95_logs.tar.gz (~2.6 GB). Contents:- logs/{task}/{seed}/{exp}/    Training output of all 95 runs (4 tasks:                  cheetah-run, quadruped-walk, pendulum-swingup,                  walker-walk; up to 10 arms; 5 seeds each).                  Each run directory contains eval.csv                  (21 evaluation points over 100k steps),                  models/final.pt (final checkpoint), and                  eval_video/ (evaluation episode videos).- run_logs/{exp}.log        Console logs of every run, including                  train/eval lines, the four [LEAP-EVAL]                  violation fields, and Lagrangian dual-variable                  (lambda) update trajectories; plus                  driver95.log (acceptance summary of all runs).- action_stats/          Deployment-measurement CSVs (400 episodes),                  source of Table 2.- analysis_out/          Learning-stage figures and tables,                  source of Tables 1 and 4.- stats_out/            Statistical test battery outputs                  (Welch t, paired t, Levene, Holm, etc.).- paper_assets/          Manuscript figures and lambda trajectory tables.- outputs/             Hydra runtime logs (training timeline).Provenance: every number in the paper can be reproduced from this archive.run_logs/*.log -> analyze95.py -> analysis_out/tables/final_performance.csv(Tables 1 and 4); models/final.pt -> collect_action_stats.py ->action_stats/action_stats_summary.csv (Table 2).All standard deviations are population standard deviations over 5 seeds (ddof=0).Code: github.com/ylu-lz/leap (analysis scripts:analyze95.py, stats_battery.py, collect_action_stats.py, make_paper_assets.py)