{
  // Data Arguments
  //################
  "dataset_HF": "c-bone/mattergen_den_ehull",
  "pretrained_tokenizer_dir": "HF-cif-tokenizer",
  "context_length": 1024,
  "dataset_streaming": false,
  // Filters
  "remove_CIFs_above_context": true,
  "remove_CIFs_with_unk": true,

    // Conditional Arguments
  //#######################
  "condition_columns": "['norm_Density (g/cm^3)', 'norm_energy_above_hull']" , 
  "n_prefix_tokens": 2,
  "n_hidden_cond": 1024,
  "cond_dropout": 0.01,
  "share_layers": false,
  "n_heads_sharing_slider": 2,
  "cond_lr": 0.0005,
  "cond_wd": 0.01,
  "activate_conditionality": "PKV",


  // Model Arguments
  //#################
  // Model Depth
  // n_positions has been tied to context_length
  "n_embd": 512,
  "n_layer": 8,
  "n_head": 8,
  // Dropout
  "residual_dropout": 0.1,
  "embedding_dropout": 0.1,
  "attention_dropout": 0.1,

  // Trainer Arguments
  //###################
  // Batching
  "train_batch_size": 32,
  "eval_batch_size": 32,
  "gradient_accumulation_steps": 1,
  "auto_find_batch_size": false,
  // Learning Rate and Optimizer
  "learning_rate": 0.000005,
  "lr_scheduler_type": "cosine_with_min_lr",
  "lr_scheduler_kwargs": {
    "min_lr_rate": 0.01
  },
  "warmup_ratio": 0.02, // 2% of training steps
  "adam_beta1": 0.9,
  "adam_beta2": 0.999,
  "grad_clip": 1.0,
  // "max_grad_norm": 1.0,
  "weight_decay": 0.01,
  // Logging
  "output_dir": "model_ckpts/mattergen_den_ehull/PKV_ft",
  "save_total_limit": 2,
  "report_to": "wandb",
  "wandb_project_folder": "dataset_size_study",
  "pretrained_model_dir": "model_ckpts/mpdb-small-base-lematerial/checkpoint-1250000",
  "eval_strategy": "steps",
  "eval_steps": 4000, // 0.5 epoch
  "logging_steps": 50, // 0.1 epoch
  "save_strategy": "steps",
  "max_steps": 400000, // 30 epochs
  // to calculate the number of steps, use the following formula:
  // steps = (number of epochs) * (number of training samples) / (batch size)
  "early_stopping_patience": 15,
  "early_stopping_threshold": 0.000005,
  // Utils
  "seed": 2,
  "data_seed": 1,
  "load_best_model_at_end": true,
  "metric_for_best_model": "eval_loss",
  "greater_is_better": false,
  "torch_compile": true,
  "fp16": true,
  "deepspeed_config": "_config_files/deepspeed_default.json",

  // CodeCarbon Arguments
  //######################
  "codecarbon": true,
  "tracker_project": "CrystaLLM-pi"
}