metadata

license: apache-2.0
pipeline_tag: translation

Model Description

Erya is a pretrained model specifically designed for translating Ancient Chinese into Modern Chinese. It utilizes an Encoder-Decoder architecture and has been trained using a combination of DMLM (Dual Masked Language Model) and DAS (Disyllabic Aligned Substitution) techniques on datasets comprising both Ancient Chinese and Modern Chinese texts. The detailed information of our work can bu found here: RUCAIBox/Erya (github.com)

Erya has not undergone fine-tuning for the machine translation task, making it possible to further enhance its translation capabilities by fine-tuning on a smaller translation dataset. The more information about Ancient Chinese and Modern Chinese can be found here: RUCAIBox/Erya-dataset · Datasets at Hugging Face

Example

from transformers import BertTokenizer, CPTForConditionalGeneration

tokenizer = BertTokenizer.from_pretrained("RUCAIBox/Erya")
model = CPTForConditionalGeneration.from_pretrained("RUCAIBox/Erya")

input_ids = tokenizer("安世字子孺，少以父任为郎。", return_tensors='pt')
input_ids.pop("token_type_ids")

pred_ids = model.generate(max_new_tokens=256, **input_ids)
print(tokenizer.batch_decode(pred_ids, skip_special_tokens=True))