{
  "abundance": {
    "min_abundance": 0.0001,
    "output_min": -2.0,
    "output_max": 2.0,
    "normalize": false,
    "num_bins": 5,
    "binning_method": "quantile",
    "fixed_thresholds": [
      0.0,
      0.0001,
      0.001,
      0.01,
      0.1
    ],
    "enable_regression": true,
    "regression_top_k_bins": 2
  },
  "model": {
    "vocab_size": 10000,
    "embedding_dim": 1536,
    "hidden_size": 320,
    "num_hidden_layers": 6,
    "num_attention_heads": 5,
    "intermediate_size": 1280,
    "hidden_dropout_prob": 0.1,
    "attention_probs_dropout_prob": 0.1,
    "layer_norm_eps": 1e-12,
    "max_position_embeddings": 2048,
    "pad_token_id": -1,
    "cls_token_id": -1,
    "mlp_dropout": 0.1
  },
  "loss": {
    "lambda_bin": 1.0,
    "lambda_regression": 0.3,
    "bin_loss_type": "focal",
    "regression_loss_type": "mse",
    "focal_alpha": 0.25,
    "focal_gamma": 2.0
  },
  "training": {
    "mask_prob": 0.15,
    "mask_ratio": [
      0.8,
      0.1,
      0.1
    ],
    "train_ratio": 0.9,
    "val_ratio": 0.1,
    "batch_size": 512,
    "num_workers": 0,
    "learning_rate": 5e-05,
    "weight_decay": 0.01,
    "warmup_steps": 1000,
    "lr_scheduler": "warmup_cosine",
    "min_lr_ratio": 0.1,
    "max_epochs": 50,
    "gradient_clip_val": 1.0,
    "precision": "bf16",
    "devices": [
      0
    ],
    "accelerator": "auto"
  },
  "qwen3_alignment": {
    "enabled": true,
    "embedding_path": "",
    "projection_dim": 4096,
    "temperature": 0.07,
    "loss_weight": 0.1
  },
  "ntv3_embedding_path": "../../embeddings/reference_ntv3_embeddings.h5",
  "data_path": "",
  "output_dir": "./outputs",
  "experiment_name": "MGM2-medium",
  "seed": 42
}