mars-aurora-openmars-random-init

Aurora architecture trained on OpenMARS from random initialization for comparison with pretrained fine-tuning.

This repository contains model-only weights exported from an OpenMARS fine-tuned Aurora checkpoint. Optimizer state and RNG state were intentionally stripped.

Source checkpoint

artifacts/openmars_runs/20260621_000248_random_init_a6000_6gpu_bs20_e10/checkpoint_step_9158.pt

Training configuration

• Model size: base
• Split manifest: splits/openmars_my28-34_train_my35_val.json
• Epochs: 13
• Batch size per GPU: 20
• Base LR: 0.0001
• New/Mars LR: 0.0001
• Resume checkpoint: artifacts/openmars_runs/20260621_000248_random_init_a6000_6gpu_bs12_e10/checkpoint_step_7760.pt

Loading

Use this repository with the mars_weather.make_mars_aurora helper from the project code. Load openmars_stats.json, construct the configured Mars Aurora model, and then load model.safetensors or pytorch_model.bin as a state dict.