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

DATA:
    task_name: eurlex57k
    dataset_name: eurlex
    dataset_config_name: null
    max_seq_length: 512
    overwrite_output_dir: true
    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/eurlex4.3k/train_split1057.jsonl
    validation_file: datasets/eurlex4.3k/test_unseen_split1057.jsonl
    test_file: datasets/eurlex4.3k/test_unseen_split1057.jsonl
    
    # validation_file: datasets/eurlex4.3k/test_unseen_hr.jsonl
    # test_file: datasets/eurlex4.3k/test_unseen_hr.jsonl
    label_max_seq_length: 128
    # descriptions_file: datasets/eurlex4.3k/heir_withdescriptions_4.3k_v1_nl_unseen.json
    # test_descriptions_file: datasets/eurlex4.3k/heir_withdescriptions_4.3k_v1_nl.json
    descriptions_file: datasets/eurlex4.3k/heir_withdescriptions_4.3k_web_nl_unseen.json
    test_descriptions_file: datasets/eurlex4.3k/heir_withdescriptions_4.3k_web_nl.json


    all_labels : datasets/eurlex4.3k/all_labels.txt
    test_labels: datasets/eurlex4.3k/unseen_labels_split1057.txt
    # test_labels: datasets/eurlex4.3k/unseen_labels.txt

    # max_descs_per_label: 5
    contrastive_learning_samples: 1500
    cl_min_positive_descs: 1
    # bm_short_file: datasets/eurlex4.3k/train_bmshort.txt
    # coil_cluster_mapping_path: bert_coil_map_dict_lemma255K_isotropic.json
    # coil_cluster_mapping_path: bert_coil_map_dict_lemma.json

MODEL:
    model_name_or_path: bert-base-uncased
    pretrained_model_path: /n/fs/scratch/pranjal/seed_experiments/ablation_eurlex_1_coil_web_seed2/checkpoint-5400/pytorch_model.bin
    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: prajjwal1/bert-small
    # label_model_name_or_path: bert-base-uncased
    # label_model_name_or_path: prajjwal1/bert-tiny
    encoder_model_type: bert
    use_custom_optimizer: adamw
    output_learning_rate: 1.e-4
    arch_type : 2
    add_label_name: false
    normalize_embeddings: false
    tie_weights: false
    coil: false
    # use_precomputed_embeddings: datasets/eurlex4.3k/heir_withdescriptions_4.3k_v1_embs_bert_9_96.npy
    token_dim: 16
    # num_frozen_layers: 9
    label_frozen_layers: 2

TRAINING:
    do_train: false
    do_eval: true
    per_device_train_batch_size: 1
    gradient_accumulation_steps: 8
    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: 10
    save_steps: 5400
    evaluation_strategy: steps
    eval_steps: 5000
    fp16: true
    fp16_opt_level: O1
    lr_scheduler_type: "linear" # defaults to 'linear'
    dataloader_num_workers: 8
    label_names: [labels]
    scenario: "unseen_labels"

    ddp_find_unused_parameters: false
    seed: -1