source_config/base.json

{
    "vocab_size": 10000,
    "max_seq_len": 2048,
    "dropout": 0.05,
    "pretrain_epochs": 1,
    "continued_pretrain_epochs": 5,
    "sft1_epochs": 5,
    "sft2_epochs": 1,
    "train_ratio": 0.95,
    "val_ratio": 0.04,
    "pretrain_learning_rate": 2e-4,
    "continued_pretrain_learning_rate": 1.5e-5,
    "sft1_learning_rate": 2e-5,
    "sft2_learning_rate": 1e-5,
    "accumulation_steps": 1,
    "pretrain_weight_decay": 0.2,
    "continued_pretrain_weight_decay": 0.1,
    "sft1_learning_weight_decay": 0.05,
    "sft2_learning_weight_decay": 0.05,
    "penalty_margin_weight": 0.1,
    "penalty_margin_detach_max": true,
    "penalty_entropy_weight": 0.001,
    "penalty_entropy_min_entropy": 0.5,
    "penalty_focal_weight": 0.05,
    "penalty_focal_gamma": 2.0
}