{
  "attention_probs_dropout_prob": 0.1,
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "intermediate_size": 2560,
  "max_position_embeddings": 512,
  "position_bucket_size": 32,
  "num_attention_heads": 12,
  "num_hidden_layers": 12,
  "vocab_size": 16384,
  "layer_norm_eps": 1e-05,
  "force_causal_mask": true,
  "classifier_dropout": 0.1,
  "classifier_layer_norm_eps": 1e-05,
  "num_labels": 2
}

Tokenizer file: tokenizer_ita_vs16384.json

Quick Usage

python

from transformers import AutoTokenizer, AutoModelForMaskedLM
model_id = 'jumelet/gptbert-ita-250steps-base'
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForMaskedLM.from_pretrained(model_id, trust_remote_code=True)
out = model(**tok('Hello world', return_tensors='pt'))

Forced Causal Attention

Causal attention is enforced during inference by applying a triangular future mask inside the remote code. This prevents the hybrid GPT-BERT layers from attending to future tokens even when a bidirectional mask is provided.

Sequence Classification

GPTBertForSequenceClassification mirrors the original GLUE classifier head for downstream fine-tuning.

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
model_id = 'jumelet/gptbert-ita-250steps-base'
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id, trust_remote_code=True)
outputs = model(**tok('This movie was great!', return_tensors='pt'))
print(outputs.logits)

Notes

Converted on 2025-10-07T01:13:25.492214+00:00
Weights are the exact trained parameters; no new layers were initialized.
Requires trust_remote_code=True due to custom architecture.