{
  "architecture": "ouro",
  "num_hidden_layers": 1,
  "num_attention_heads": 2,
  "loops": 4,
  "optimizer": "muon",
  "learning_rate": 0.003,
  "batch_size": 32,
  "eval_batch_size": 32,
  "max_steps": 4000,
  "eval_steps": 250,
  "seed": 456,
  "data_seed": 42,
  "hidden_size": 32,
  "intermediate_size": 48,
  "model_overrides": {
    "value_residual_mix": 0.5,
    "num_key_value_heads": 1
  }
}