bert-base-turkish-128k-cased-finetuned_lr-2e-05_epochs-3

from transformers import BertTokenizer, BertForQuestionAnswering
import torch

# Cargar el modelo pre-entrenado
model_name = 'husnu/bert-base-turkish-128k-cased-finetuned_lr-2e-05_epochs-3'
tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForQuestionAnswering.from_pretrained(model_name)

# Texto de ejemplo
question = '¿Cuál es la capital de Turquía?'
text = 'La capital de Turquía es Ankara.'

# Tokenizar input
input_ids = tokenizer.encode(question, text)

# Obtener el ID del token del separador
sep_index = input_ids.index(tokenizer.sep_token_id)

# Mensajes de segmentación
segment_ids = [0] * (sep_index + 1) + [1] * (len(input_ids) - sep_index - 1)

# Convertir a tensores de PyTorch
input_tensors = torch.tensor([input_ids])
segment_tensors = torch.tensor([segment_ids])

# Realizar inferencia
with torch.no_grad():
    outputs = model(input_tensors, token_type_ids=segment_tensors)
    start_scores, end_scores = outputs.start_logits, outputs.end_logits

# Encontrar inicio y fin de la respuesta
all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
start_index = torch.argmax(start_scores)
end_index = torch.argmax(end_scores) + 1
answer = tokenizer.convert_tokens_to_string(all_tokens[start_index:end_index])

print(f'Respuesta: {answer}')