tuning_config.json

{
  "paths" : {
    "main_path" : "data",
    "train_path" : "%main_path%/eltec_ud_ner.jsonl",
    "dev_path" : "%main_path%/dev_SrpKor4Tagging.jsonl",
    "tokenizer_path" :"%main_path%/tokenizer.json",
    "model_folder" : "saved/tune-saved",
    "pretrained" : "C:\\Users\\mihailo\\OneDrive - Faculty of Mining and Geology\\resursi\\bert modeli\\bertovic-base"
  },

  "model-options": {
    "model_type" : "roberta-base",
    "tuning_task": "token-classification",
    "token_col": "token",
    "label_col": "ner"
  },

  "training-options": {
    "num_train_epochs" : 2,
    "per_device_train_batch_size" : 2,
    "per_device_eval_batch_size" : 2,
    "learning_rate" : 0.00005,
    "weight_decay" : 0,
    "warmup_steps" : 2000,

    "save_total_limit" : 1,
    "load_best_model_at_end" : true,
    "overwrite_output_dir" : true,
    "evaluation_strategy" : "epoch",
    "output_dir" : "saved/tune-saved",
    "save_strategy" : "epoch"
  },

  "tokenizer_training": {
    "path" : "C:/gpt2/tokenizer/",
    "size": 49152,
    "freq": 2,
    "special_tokens": ["<s>", "<pad>", "</s>", "<unk>", "<mask>"],
    "unk_token": "<UNK>",
    "type": "BPE"
  }
}