config.toml

# =============================================================================
# Before you start changing anything here, read the comments.
# All of them can be found below in the "DEFAULT" section

[DEFAULT]

# The directory that contains extracted files of everything you've downloaded.
data_dir = "data"

# Train, dev and test jsonlines
train_data = "data/english_train_head.jsonlines"
dev_data = "data/english_development_head.jsonlines"
test_data = "data/english_test_head.jsonlines"

# The device where everything is to be placed. "cuda:N"/"cpu" are supported.
device = "cuda:0"


# Bert settings ======================

# Base bert model architecture and tokenizer
bert_model = "bert-large-cased"

# Controls max length of sequences passed through bert to obtain its
# contextual embeddings
# Must be less than or equal to 512
bert_window_size = 512


# General model settings =============

# Controls the dimensionality of feature embeddings
embedding_size = 20

# Controls the dimensionality of distance embeddings used by SpanPredictor
sp_embedding_size = 64

# Controls the number of spans for which anaphoricity can be scores in one
# batch. Only affects final scoring; mention extraction and rough scoring
# are less memory intensive, so they are always done in just one batch.
a_scoring_batch_size = 512

# AnaphoricityScorer FFNN parameters
hidden_size = 1024
n_hidden_layers = 1


# Mention extraction settings ========

# Mention extractor will check spans up to max_span_len words
# The default value is chosen to be big enough to hold any dev data span
max_span_len = 64


# Pruning settings ===================

# Controls how many pairs should be preserved per mention
# after applying rough scoring.
rough_k = 50


# Training settings ==================

# Controls whether to fine-tune bert_model
bert_finetune = true

# Controls the dropout rate throughout all models
dropout_rate = 0.3

# Bert learning rate (only used if bert_finetune is set)
bert_learning_rate = 1e-5

# Task learning rate
learning_rate = 3e-4

# For how many epochs the training is done
train_epochs = 20

# Controls the weight of binary cross entropy loss added to nlml loss
bce_loss_weight = 0.5

# The directory that will contain conll prediction files
conll_log_dir = "data/conll_logs"

# =============================================================================
# Extra keyword arguments to be passed to bert tokenizers of specified models
[DEFAULT.tokenizer_kwargs]
    [DEFAULT.tokenizer_kwargs.roberta-large]
        "add_prefix_space" = true

    [DEFAULT.tokenizer_kwargs.spanbert-large-cased]
        "do_lower_case" = false

    [DEFAULT.tokenizer_kwargs.bert-large-cased]
        "do_lower_case" = false

# =============================================================================
# The sections listed here do not need to make use of all config variables
# If a variable is omitted, its default value will be used instead

[roberta]
bert_model = "roberta-large"

[roberta_no_finetune]
bert_model = "roberta-large"
bert_finetune = false

[roberta_no_bce]
bert_model = "roberta-large"
bce_loss_weight = 0.0

[spanbert]
bert_model = "SpanBERT/spanbert-large-cased"

[spanbert_no_bce]
bert_model = "SpanBERT/spanbert-large-cased"
bce_loss_weight = 0.0

[bert]
bert_model = "bert-large-cased"

[longformer]
bert_model = "allenai/longformer-large-4096"
bert_window_size = 2048

[debug]
bert_window_size = 384
bert_finetune = false
device = "cpu:0"

[debug_gpu]
bert_window_size = 384
bert_finetune = false