GLiNER

1
# Model Configuration
2
model_name: microsoft/deberta-v3-small  # Hugging Face model
3
name: "large"
4
max_width: 12
5
hidden_size: 768
6
dropout: 0.4
7
fine_tune: true
8
subtoken_pooling: first
9
span_mode: markerV0
10

11
# Training Parameters
12
num_steps: 30000
13
train_batch_size: 3
14
eval_every: 5000
15
warmup_ratio: 0.1
16

17
# Learning Rate Configuration
18
lr_encoder: 1e-5
19
lr_others: 5e-5
20

21
# Directory Paths
22
root_dir: gliner_logs
23
train_data: "data.json" # see https://github.com/urchade/GLiNER/tree/main/data
24
val_data_dir: "none"
25
# "NER_datasets": val data from the paper can be obtained from "https://drive.google.com/file/d/1T-5IbocGka35I7X3CE6yKe5N_Xg2lVKT/view"
26

27
# Pretrained Model Path
28
# Use "none" if no pretrained model is being used
29
prev_path: "none"
30

31
# Advanced Training Settings
32
size_sup: -1
33
max_types: 25
34
shuffle_types: true
35
random_drop: true
36
max_neg_type_ratio: 1
37
max_len: 384
38
freeze_token_rep: false
39