hyper_config.json

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