{
  "embedding_key": "model.embed_tokens",
  "max_iter": 4,
  "iter_decider": "Qwen3MLPIterDecider",
  "input_updater": "Qwen3MLPUpdater",
  "train_loss": "CombinedLoss",
  "eval_loss": "NextTokenPredLoss",
  "eval_iter_decider": "iter_decider",
  "iter_label_generator": "DynamicIterLabelGenerator",
  "iter_attention_mode": "duo",
  "attn_implementation": "sdpa",
  "iter_attention_impl": "triton",
  "weighted_hidden_method": "stop_prob_mix",
  "memory_lean_fp32_reductions": false,
  "detach_cross_iter_kv": false,
  "iter_decider_kwargs": {
    "hidden_size": 2048,
    "intermediate_size": 2048,
    "threshold": 0.5,
    "topk": 2048
  },
  "input_updater_kwargs": {
    "hidden_size": 2048,
    "intermediate_size": 2048
  },
  "train_loss_kwargs": {
    "losses": [
      {
        "name": "NextTokenPredLoss",
        "weight": 1.0
      },
      {
        "name": "IterDeciderLoss",
        "weight": 0.05,
        "kwargs": {
          "dyn_pos_weight": true,
          "cost_sensitive": true,
          "cost_aware": false
        }
      }
    ]
  },
  "eval_loss_kwargs": {},
  "eval_iter_decider_kwargs": {},
  "iter_label_generator_kwargs": {
    "strategy": "posterior_ce",
    "abs_improv_coverage": 0.99
  }
}