EXP_NAME: "semsup_descs_100ep_newds_cosine"             
EXP_DESC: "SemSup Descriptions ran for 100 epochs"

DATA:
    task_name: amazon13k
    dataset_name: amazon13k
    dataset_config_name: null
    max_seq_length: 128
    overwrite_output_dir: true # Set to false, if using one_hour_job
    overwrite_cache: false
    pad_to_max_length: true
    load_from_local: true
    max_train_samples: null
    max_eval_samples: null
    max_predict_samples: null
    train_file: datasets/Amzn13K/train_split1668_random.jsonl
    validation_file: datasets/Amzn13K/test_unseen_split6500_2.jsonl
    test_file: datasets/Amzn13K/test_unseen_split6500_2.jsonl
    label_max_seq_length: 32
    # descriptions_file: datasets/Amzn13K/amzn_curie_descsriptions.json
    descriptions_file: datasets/Amzn13K/amzn_descs_refined_v3_v3.json
    all_labels : datasets/Amzn13K/all_labels.txt
    test_labels: datasets/Amzn13K/unseen_labels_split6500_2.txt

    # max_descs_per_label: 10
    # contrastive_learning_samples: 5000
    # cl_min_positive_descs: 1
    # bm_short_file: datasets/eurlex4.3k/train_bmshort.txt
    # ignore_pos_labels_file: datasets/Amzn13K/ignore_train_split1668_random_fs50.txt

MODEL:
    model_name_or_path: bert-base-uncased
    config_name: null
    tokenizer_name: null
    cache_dir: null
    use_fast_tokenizer: true
    model_revision: main
    use_auth_token: false
    ignore_mismatched_sizes: false
    negative_sampling: "none"
    semsup: true
    # label_model_name_or_path: bert-base-uncased # prajjwal1/bert-small
    label_model_name_or_path: prajjwal1/bert-small
    encoder_model_type: bert
    use_custom_optimizer: adamw
    output_learning_rate: 1.e-4
    arch_type : 2
    add_label_name: true
    normalize_embeddings: false
    tie_weights: false
    coil: true
    colbert: false
    # use_precomputed_embeddings: datasets/eurlex4.3k/heir_withdescriptions_4.3k_v1_embs_bert_9_96.npy
    token_dim: 16
    label_frozen_layers: 2

TRAINING:
    do_train: true
    do_eval: true
    do_predict: false
    per_device_train_batch_size: 1
    gradient_accumulation_steps: 4
    per_device_eval_batch_size: 1
    learning_rate: 5.e-5 # Will point to input encoder lr, if user_custom_optimizer is False
    num_train_epochs: 20
    save_steps: 5000
    evaluation_strategy: steps
    eval_steps: 2000
    fp16: true
    fp16_opt_level: O1
    lr_scheduler_type: "linear" # defaults to 'linear'
    dataloader_num_workers: 4
    label_names: [labels]
    scenario: "unseen_labels"

    ddp_find_unused_parameters: false
    max_eval_samples: 15000
    max_train_samples: 30000

    # ignore_data_skip: true
    # one_hour_job: true