sft_config.json

{
  "name": "dmt-multilingual-50m-translation-sft",
  "base_model_dir": "outputs/dmt-multilingual-50m/final",
  "data_dir": "data/sft_tokenized",
  "tokenizer_path": "tokenizer/tokenizer.json",
  "output_dir": "outputs/dmt-multilingual-50m-translation-sft",
  "training": {
    "micro_batch_size": 64,
    "eval_micro_batch_size": 64,
    "gradient_accumulation_steps": 2,
    "epochs": 1,
    "max_steps": 0,
    "learning_rate": 5e-05,
    "min_learning_rate": 5e-06,
    "weight_decay": 0.01,
    "warmup_ratio": 0.03,
    "gradient_clip": 1.0,
    "num_workers": 8,
    "pad_to_multiple_of": 8,
    "log_interval": 20,
    "eval_interval": 500,
    "eval_batches": 100,
    "save_interval": 2000,
    "seed": 42,
    "compile": false,
    "gradient_checkpointing": false
  }
}