openprompt

1
dataset:
2
  name: super_glue.copa
3
  path:  # dataset in huggingface doesn't need path
4

5
plm:
6
  model_name: bert
7
  model_path: bert-large-cased
8
  optimize:
9
    freeze_para: True
10
    lr: 1.0e-5
11
    weight_decay: 0.0
12
    scheduler:
13
      type: 
14
      num_warmup_steps: 500
15

16
dataloader:
17
  max_seq_length: 384 # max_seq_length 
18
  decoder_max_length: 3 # the decoder max length to truncate decoder input sequence
19
                    # if it is an encoder-decoder architecture. Note that it's not equavalent
20
                    # to generation.max_length which is used merely in the generation phase.
21
  truncate_method: "head" # choosing from balanced, head, tail
22
  decode_from_pad: false
23

24
train:
25
  batch_size: 4
26
  gradient_accumulation_steps: 1
27
  max_grad_norm: 1.0
28
  num_epochs:
29
  num_training_steps: 30000
30
  teacher_forcing: false
31

32

33
test:
34
  batch_size: 16
35

36
dev:
37
  batch_size: 16
38

39

40
template: soft_template
41
verbalizer: contextual_verbalizer
42

43

44

45
soft_template:
46
  choice: 0
47
  file_path: scripts/SuperGLUE/COPA/soft_template.txt
48
  num_tokens: 20
49
  initialize_from_vocab: true
50
  random_range: 0.5
51
  optimize: 
52
    name: AdamW
53
    lr: 0.03
54
    adam_epsilon: 1.0e-8
55
    scheduler:
56
      num_warmup_steps: 500
57

58

59
contextual_verbalizer:
60
  
61
environment:
62
  num_gpus: 1
63
  cuda_visible_devices:
64
  local_rank: 0 
65

66
learning_setting: full #few_shot
67

68
# few_shot:
69
#   parent_config: learning_setting
70
#   few_shot_sampling: sampling_from_train
71
  
72
# sampling_from_train:
73
#   parent_config: few_shot_sampling
74
#   num_examples_per_label: 100
75
#   also_sample_dev: True
76
#   num_examples_per_label_dev: 100
77
#   seed:
78
#     - 123
79