{ "training_step": 82697, "base_model": "Qwen/Qwen2.5-3B-Instruct", "latent_depth_trained": 4, "halt_threshold_default": 0.7 }