jzhoubu/vdr-msmarco

biencoder:
  alias: vdr
  sequence_length: 512
  dropout: 0.1
  shared_encoder: false
  semiparametric: true
  device: null
  encoder_q:
    model_id: bert-base-uncased
    max_seq_len: 256
    pretrained: true
    norm: false
    shift_vocab_num: 1000
    prefix: encoder_q.
  encoder_p:
    model_id: bert-base-uncased
    max_seq_len: 256
    pretrained: true
    norm: false
    shift_vocab_num: 1000
    prefix: encoder_p.
train_datasets:
- marco_train_bm25
dev_datasets:
- marco_dev_30neg
datastore: null
output_dir: null
train_sampling_rates: null
batch_size: 64
num_train_epochs: 20
num_warmup_epochs: 1
num_save_epochs: 1
num_eval_epochs: 1
hard_negatives: 1
other_negatives: 0
ret_negatives: 0
train_insert_title: false
valid_insert_title: false
adam_eps: 1.0e-08
adam_betas: (0.9, 0.999)
learning_rate: 2.0e-05
max_grad_norm: 2.0
log_batch_step: 100
train_rolling_loss_step: 100
weight_decay: 0.0
sym_loss: true
do_lower_case: true
seed: 12345
checkpoint_file_name: vdr
save_every_epoch: true
model_file: null
local_rank: -1
local-rank: 2
device: null
distributed_world_size: null
distributed_port: null
no_cuda: false
n_gpu: null
fp16: true