Child R6 class for creation and training of ModernBERT transformers

This class has the following methods:

• create: creates a new transformer based on ModernBERT.

• train: trains and fine-tunes a ModernBERT model.

Note

This model uses a WordPiece tokenizer like BERT and can be trained with whole word masking. The transformer library may display a warning, which can be ignored.

Create

New models can be created using the .AIFEModernBertTransformer$create method.

Train

To train the model, pass the directory of the model to the method .AIFEModernBertTransformer$train.

Pre-Trained models that can be fine-tuned using this method are available at https://huggingface.co/.

The model is trained using dynamic masking, as opposed to the original paper, which used static masking.

References

Devlin, J., Chang, M.‑W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In J. Burstein, C. Doran, & T. Solorio (Eds.), Proceedings of the 2019 Conference of the North (pp. 4171–4186). Association for Computational Linguistics. doi:10.18653/v1/N19-1423

Hugging Face documentation

• https://huggingface.co/docs/transformers/model_doc/bert

• https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertForMaskedLM

• https://huggingface.co/docs/transformers/model_doc/bert#transformers.TFBertForMaskedLM

See also

Other Transformers for developers: .AIFETrConfig, .AIFETrModel, .AIFETrModelMLM, .AIFETrObj, .AIFETrTokenizer, .aife_transformer.check_type(), aife_transformer.load_model(), aife_transformer.load_model_config(), aife_transformer.load_model_mlm(), aife_transformer.load_tokenizer()

Super class

aifeducation::.AIFEBaseTransformer -> .AIFEModernBertTransformer

Methods

Public methods

• .AIFEModernBertTransformer$new()

• .AIFEModernBertTransformer$create()

• .AIFEModernBertTransformer$train()

• .AIFEModernBertTransformer$clone()

Inherited methods

• aifeducation::.AIFEBaseTransformer$init_transformer()
• aifeducation::.AIFEBaseTransformer$set_SFC_calculate_vocab()
• aifeducation::.AIFEBaseTransformer$set_SFC_check_max_pos_emb()
• aifeducation::.AIFEBaseTransformer$set_SFC_create_final_tokenizer()
• aifeducation::.AIFEBaseTransformer$set_SFC_create_tokenizer_draft()
• aifeducation::.AIFEBaseTransformer$set_SFC_create_transformer_model()
• aifeducation::.AIFEBaseTransformer$set_SFC_save_tokenizer_draft()
• aifeducation::.AIFEBaseTransformer$set_SFT_create_data_collator()
• aifeducation::.AIFEBaseTransformer$set_SFT_cuda_empty_cache()
• aifeducation::.AIFEBaseTransformer$set_SFT_load_existing_model()
• aifeducation::.AIFEBaseTransformer$set_model_param()
• aifeducation::.AIFEBaseTransformer$set_model_temp()
• aifeducation::.AIFEBaseTransformer$set_required_SFC()
• aifeducation::.AIFEBaseTransformer$set_title()

---

Method new()

Creates a new transformer based on ModernBERT and sets the title.

Usage

.AIFEModernBertTransformer$new(init_trace = TRUE)

Arguments

init_trace

bool option to show prints. If TRUE (by default) - messages will be shown, otherwise (FALSE) - hidden.

Returns

This method returns nothing.

---

Method create()

This method creates a transformer configuration based on the ModernBERT base architecture and a vocabulary based on WordPiece by using the python libraries transformers and tokenizers.

This method adds the following 'dependent' parameters to the base class's inherited params list:

• vocab_do_lower_case

• num_hidden_layer

Usage

.AIFEModernBertTransformer$create(
  model_dir,
  text_dataset,
  vocab_size = 30522,
  vocab_do_lower_case = FALSE,
  max_position_embeddings = 512,
  hidden_size = 768,
  num_hidden_layer = 12,
  num_attention_heads = 12,
  intermediate_size = 3072,
  hidden_act = "GELU",
  hidden_dropout_prob = 0.1,
  attention_probs_dropout_prob = 0.1,
  sustain_track = FALSE,
  sustain_iso_code = NULL,
  sustain_region = NULL,
  sustain_interval = 15,
  trace = TRUE,
  pytorch_safetensors = TRUE,
  log_dir = NULL,
  log_write_interval = 2
)

Arguments

model_dir

string Path to the directory where the model should be saved. Allowed values: any

text_dataset

LargeDataSetForText LargeDataSetForText Object storing textual data.

vocab_size

int Size of the vocabulary. Allowed values: 1000 <= x <= 5e+05

vocab_do_lower_case

bool TRUE if all words/tokens should be lower case.

max_position_embeddings

int Number of maximum position embeddings. This parameter also determines the maximum length of a sequence which can be processed with the model. Allowed values: 10 <= x <= 4048

hidden_size

int Number of neurons in each layer. This parameter determines the dimensionality of the resulting text embedding. Allowed values: 1 <= x <= 2048

num_hidden_layer

int Number of hidden layers. Allowed values: 1 <= x

num_attention_heads

int determining the number of attention heads for a self-attention layer. Only relevant if attention_type='multihead' Allowed values: 0 <= x

intermediate_size

int determining the size of the projection layer within a each transformer encoder. Allowed values: 1 <= x

hidden_act

string Name of the activation function. Allowed values: 'gelu', 'relu', 'silu', 'gelu_new'

hidden_dropout_prob

double Ratio of dropout. Allowed values: 0 <= x <= 0.6

attention_probs_dropout_prob

double Ratio of dropout for attention probabilities. Allowed values: 0 <= x <= 0.6

sustain_track

bool If TRUE energy consumption is tracked during training via the python library 'codecarbon'.

sustain_iso_code

string ISO code (Alpha-3-Code) for the country. This variable must be set if sustainability should be tracked. A list can be found on Wikipedia: https://en.wikipedia.org/wiki/List_of_ISO_3166_country_codes. Allowed values: any

sustain_region

string Region within a country. Only available for USA and Canada See the documentation of codecarbon for more information. https://mlco2.github.io/codecarbon/parameters.html Allowed values: any

sustain_interval

int Interval in seconds for measuring power usage. Allowed values: 1 <= x

trace

bool TRUE if information about the estimation phase should be printed to the console.

pytorch_safetensors

bool

• TRUE: a 'pytorch' model is saved in safetensors format.

• FALSE (or 'safetensors' is not available): model is saved in the standard pytorch format (.bin).

log_dir

string Path to the directory where the log files should be saved. If no logging is desired set this argument to NULL. Allowed values: any

log_write_interval

int Time in seconds determining the interval in which the logger should try to update the log files. Only relevant if log_dir is not NULL. Allowed values: 1 <= x

Returns

This method does not return an object. Instead, it saves the configuration and vocabulary of the new model to disk.

---

Method train()

This method can be used to train or fine-tune a transformer based on ModernBERT architecture with the help of the python libraries transformers, datasets, and tokenizers.

Usage

.AIFEModernBertTransformer$train(
  output_dir,
  model_dir_path,
  text_dataset,
  p_mask = 0.15,
  whole_word = TRUE,
  val_size = 0.1,
  n_epoch = 1,
  batch_size = 12,
  chunk_size = 250,
  full_sequences_only = FALSE,
  min_seq_len = 50,
  learning_rate = 0.003,
  sustain_track = FALSE,
  sustain_iso_code = NULL,
  sustain_region = NULL,
  sustain_interval = 15,
  trace = TRUE,
  pytorch_trace = 1,
  pytorch_safetensors = TRUE,
  log_dir = NULL,
  log_write_interval = 2
)

Arguments

output_dir

string Path to the directory where the model should be saved. Allowed values: any

model_dir_path

string Path to the directory where the original model is stored. Allowed values: any

text_dataset

LargeDataSetForText LargeDataSetForText Object storing textual data.

p_mask

double Ratio that determines the number of words/tokens used for masking. Allowed values: 0 < x < 1

whole_word

bool * TRUE: whole word masking should be applied.

• FALSE: token masking is used.

val_size

double between 0 and 1, indicating the proportion of cases which should be used for the validation sample during the estimation of the model. The remaining cases are part of the training data. Allowed values: 0 < x < 1

n_epoch

int Number of training epochs. Allowed values: 1 <= x

batch_size

int Size of the batches for training. Allowed values: 1 <= x

chunk_size

int Maximum length of every sequence. Must be equal or less the global maximum size allowed by the model. Allowed values: 100 <= x

full_sequences_only

bool TRUE for using only chunks with a sequence length equal to chunk_size.

min_seq_len

bool TRUE for using only chunks with a sequence length equal to chunk_size.

learning_rate

double Initial learning rate for the training. Allowed values: 0 < x <= 1

sustain_track

bool If TRUE energy consumption is tracked during training via the python library 'codecarbon'.

sustain_iso_code

string ISO code (Alpha-3-Code) for the country. This variable must be set if sustainability should be tracked. A list can be found on Wikipedia: https://en.wikipedia.org/wiki/List_of_ISO_3166_country_codes. Allowed values: any

sustain_region

string Region within a country. Only available for USA and Canada See the documentation of codecarbon for more information. https://mlco2.github.io/codecarbon/parameters.html Allowed values: any

sustain_interval

int Interval in seconds for measuring power usage. Allowed values: 1 <= x

trace

bool TRUE if information about the estimation phase should be printed to the console.

pytorch_trace

int ml_trace=0 does not print any information about the training process from pytorch on the console. Allowed values: 0 <= x <= 1

pytorch_safetensors

bool

• TRUE: a 'pytorch' model is saved in safetensors format.

• FALSE (or 'safetensors' is not available): model is saved in the standard pytorch format (.bin).

log_dir

string Path to the directory where the log files should be saved. If no logging is desired set this argument to NULL. Allowed values: any

log_write_interval

int Time in seconds determining the interval in which the logger should try to update the log files. Only relevant if log_dir is not NULL. Allowed values: 1 <= x

Returns

This method does not return an object. Instead the trained or fine-tuned model is saved to disk.

---

Method clone()

The objects of this class are cloneable with this method.

Usage

.AIFEModernBertTransformer$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.