{
  # parallelism settings
  "pipe-parallel-size": 1,
  "model-parallel-size": 4,

  # model settings
  "num-layers": 16,
  "hidden-size": 4096,
  "num-attention-heads": 32,
  "seq-length": 4096,
  "max-position-embeddings": 4096,

  # architecture design
  "norm": "layernorm",
  "pos-emb": "rotary",
  "rotary_pct": 0.25,
  "activation": "gelu",
  "no-weight-tying": true,
  "gpt_j_residual": true,
  "output_layer_parallelism": "column",

  # init methods
  "init_method": "small_init",
  "output_layer_init_method": "wang_init",

  # fused ops
  "scaled-upper-triang-masked-softmax-fusion": true,
  "bias-gelu-fusion": true,
  "attention-config": [[["flash"], 16]],

  # optimizer settings
  "optimizer": {
    "type": "Adam",
    "params": {
      "lr": 1.6e-4,
      "betas": [0.9, 0.9999],
      "eps": 1.0e-6
    },
  },
  "min_lr": 1.6e-5,

  # for all zero_optimization options, see https://www.deepspeed.ai/docs/config-json/#zero-optimizations-for-fp16-training
  "zero_optimization": {
    "stage": 1,
    "allgather_partitions": true,
    "allgather_bucket_size": 1260000000,
    "overlap_comm": true,
    "reduce_scatter": true,
    "reduce_bucket_size": 1260000000,
    "contiguous_gradients": true,
    "cpu_offload": false,
  },

  # batch / data settings
  "train_micro_batch_size_per_gpu": 32,
  "gradient_accumulation_steps": 1,
  "eval_batch_size": 2,
  "data-impl": "mmap",

  # activation checkpointing
  "checkpoint-activations": true,
  "checkpoint-num-layers": 1,
  "partition-activations": true,
  "synchronize-each-layer": true,

  # regularization
  "gradient_clipping": 1.0,
  "weight-decay": 0.1,
  "hidden-dropout": 0,
  "attention-dropout": 0,

  # precision settings
  "fp16": {
    "fp16": true,
    "enabled": true,
    "loss_scale_window": 1000,
    "initial_scale_power": 12,
    "hysteresis": 20,
    "min_loss_scale": 1,
  },

  # misc. training settings
  "train-iters": 180000,
  "lr-decay-iters": 180000,
  "distributed-backend": "nccl",
  "lr-decay-style": "cosine",
  "warmup": 0.01,
  "checkpoint-factor": 1000,
  # 1 more than checkpoint-factor to avoid skipping evals if `evaluate` fails
  "eval-interval": 1001, 
  "eval-iters": 10,
  "eval_tasks": ["piqa", "sciq", "lambada_openai"], 

  # checkpoint settings
  "iteration": 84000,
  "save": "PATH_TO_SAVE_THE_MODEL",
  "load": "PATH_TO_LOAD_THE_MODEL",

  # data settings
  "train-data-paths": [],
  "train-data-weights": [0.03, 0.02, 3.5, 13.87, 30.88, 0.34, 0.03, 0.1, 0.01, 0.5, 0.25, 0.25, 0.11, 1, 0.1, 0.5, 0.6, 0.19, 0.05, 2, 3.5, 4.15, 5.75, 3.17, 3.44, 3.49, 3, 6, 0.02, 2, 0.01, 3.95, 0.05, 1.09, 6.05],
  "valid-data-paths": [],
  "valid-data-weights": [0.03, 0.02, 3.5, 13.87, 30.88, 0.34, 0.03, 0.1, 0.01, 0.5, 0.25, 0.25, 0.11, 1, 0.1, 0.5, 0.6, 0.19, 0.05, 2, 3.5, 4.15, 5.75, 3.17, 3.44, 3.49, 3, 6, 0.02, 2, 0.01, 3.95, 0.05, 1.09, 6.05],
  "test-data-paths": [],
  "test-data-weights": [0.03, 0.02, 3.5, 13.87, 30.88, 0.34, 0.03, 0.1, 0.01, 0.5, 0.25, 0.25, 0.11, 1, 0.1, 0.5, 0.6, 0.19, 0.05, 2, 3.5, 4.15, 5.75, 3.17, 3.44, 3.49, 3, 6, 0.02, 2, 0.01, 3.95, 0.05, 1.09, 6.05],

  # tokenizer settings
  "tokenizer-type": "HFTokenizer",
  "vocab-file": "/pile/20B_tokenizer.json",

  # log settings
  "log-interval": 10,
  "steps_per_print": 10,
  "wall_clock_breakdown": true,
  "log-grad-norm": true,

  "use_wandb": true,
  "wandb_host": "",
  "wandb_team": "",
  "wandb_project": "",
  "wandb_group": "",
  "wandb_name": "",

  # multi-node launcher
  "launcher": "slurm",
  "deepspeed_slurm": true
}