Mbart Large 50 Verbalization

🚀 mbart-large-50-verbalization模型卡片

mbart-large-50-verbalization 是 facebook/mbart-large-50 模型的微调版本，专为将乌克兰语文本进行口语化处理而设计，以便用于文本转语音（TTS）系统。该模型旨在将数字和日期等结构化数据转换为乌克兰语的完整扩展文本表示。

🚀 快速开始

普通使用

from transformers import MBartForConditionalGeneration, AutoTokenizer
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

model_name = "skypro1111/mbart-large-50-verbalization"

model = MBartForConditionalGeneration.from_pretrained(
        model_name,
        low_cpu_mem_usage=True,
        device_map=device,
)
model.eval()
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.src_lang = "uk_XX"
tokenizer.tgt_lang = "uk_XX"

input_text = "<verbalization>:Цей додаток вийде 15.06.2025."

encoded_input = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=1024).to(device)
output_ids = model.generate(**encoded_input, max_length=1024, num_beams=5, early_stopping=True)
output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)

print(output_text)

ONNX使用

poetry new verbalizer
rm -rf verbalizer/tests/ verbalizer/verbalizer/ verbalizer/README.md 
cd verbalizer/
poetry shell
wget https://huggingface.co/skypro1111/mbart-large-50-verbalization/resolve/main/onnx/infer_onnx_hf.py
poetry add transformers huggingface_hub onnxruntime-gpu torch
python infer_onnx_hf.py

import onnxruntime
import numpy as np
from transformers import AutoTokenizer
import time
import os
from huggingface_hub import hf_hub_download

model_name = "skypro1111/mbart-large-50-verbalization"


def download_model_from_hf(repo_id=model_name, model_dir="./"):
    """Download ONNX models from HuggingFace Hub."""
    
    files = ["onnx/encoder_model.onnx", "onnx/decoder_model.onnx", "onnx/decoder_model.onnx_data"]

    for file in files:
        hf_hub_download(
            repo_id=repo_id,
            filename=file,
            local_dir=model_dir,
        )
    
    return files

def create_onnx_session(model_path, use_gpu=True):
    """Create an ONNX inference session."""
    # Session options
    session_options = onnxruntime.SessionOptions()
    session_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
    session_options.enable_mem_pattern = True
    session_options.enable_mem_reuse = True
    session_options.intra_op_num_threads = 8
    session_options.log_severity_level = 1
    
    cuda_provider_options = {
        'device_id': 0,
        'arena_extend_strategy': 'kSameAsRequested',
        'gpu_mem_limit': 0,  # 0 means no limit
        'cudnn_conv_algo_search': 'DEFAULT',
        'do_copy_in_default_stream': True,
    }
    
    print(f"Available providers: {onnxruntime.get_available_providers()}")
    if use_gpu and 'CUDAExecutionProvider' in onnxruntime.get_available_providers():
        providers = [('CUDAExecutionProvider', cuda_provider_options)]
        print("Using CUDA for inference")
    else:
        providers = ['CPUExecutionProvider']
        print("Using CPU for inference")
    
    session = onnxruntime.InferenceSession(
        model_path,
        providers=providers,
        sess_options=session_options
    )
    
    return session

def generate_text(text, tokenizer, encoder_session, decoder_session, max_length=128):
    """Generate text for a single input."""
    # Prepare input
    inputs = tokenizer(text, return_tensors="np", padding=True, truncation=True, max_length=512)
    input_ids = inputs["input_ids"].astype(np.int64)
    attention_mask = inputs["attention_mask"].astype(np.int64)
    
    # Run encoder
    encoder_outputs = encoder_session.run(
        output_names=["last_hidden_state"],
        input_feed={
            "input_ids": input_ids,
            "attention_mask": attention_mask,
        }
    )[0]
    
    # Initialize decoder input
    decoder_input_ids = np.array([[tokenizer.pad_token_id]], dtype=np.int64)
    
    # Generate sequence
    for _ in range(max_length):
        # Run decoder
        decoder_outputs = decoder_session.run(
            output_names=["logits"],
            input_feed={
                "input_ids": decoder_input_ids,
                "encoder_hidden_states": encoder_outputs,
                "encoder_attention_mask": attention_mask,
            }
        )[0]
        
        # Get next token
        next_token = decoder_outputs[:, -1:].argmax(axis=-1)
        decoder_input_ids = np.concatenate([decoder_input_ids, next_token], axis=-1)
        
        # Check if sequence is complete
        if tokenizer.eos_token_id in decoder_input_ids[0]:
            break
    
    # Decode sequence
    output_text = tokenizer.decode(decoder_input_ids[0], skip_special_tokens=True)
    return output_text

def main():
    # Print available providers
    print("Available providers:", onnxruntime.get_available_providers())
    
    # Download models from HuggingFace
    print("\nDownloading models from HuggingFace...")
    encoder_path, decoder_path, _ = download_model_from_hf()
    
    # Load tokenizer and models
    print("\nLoading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    tokenizer.src_lang = "uk_UA"
    tokenizer.tgt_lang = "uk_UA"
    
    # Create ONNX sessions
    print("\nLoading encoder...")
    encoder_session = create_onnx_session(encoder_path)
    print("\nLoading decoder...")
    decoder_session = create_onnx_session(decoder_path)
    
    # Test examples
    test_inputs = [
        "мій телефон 0979456822",
        "квартира площею 11 тис кв м.",
        "Пропонували хабар у 1 млрд грн.",
        "1 2 3 4 5 6 7 8 9 10.",
        "Крім того, парламентарій володіє шістьма ділянками землі (дві площею 25000 кв м, дві по 15000 кв м та дві по 10000 кв м) розташованими в Сосновій Балці Луганської області.",
        "Підписуючи цей документ у 2003 році, голови Росії та України мали намір зміцнити співпрацю та сприяти розширенню двосторонніх відносин.",
        "Очікується, що цей застосунок буде запущено 22.08.2025.",
        "За інформацією від Державної служби з надзвичайних ситуацій станом на 7 ранку 15 липня.",
    ]
    
    print("\nWarming up...")
    _ = generate_text(test_inputs[0], tokenizer, encoder_session, decoder_session)
    
    print("\nRunning inference...")
    for text in test_inputs:
        print(f"\nInput: {text}")
        t = time.time()
        output = generate_text(text, tokenizer, encoder_session, decoder_session)
        print(f"Output: {output}")
        print(f"Time: {time.time() - t:.2f} seconds")

if __name__ == "__main__":
    main() 

✨ 主要特性

• 基于强大的 facebook/mbart-large-50 架构，在多语言翻译和文本生成任务中表现出色。
• 专门针对乌克兰语文本进行微调，能将结构化数据转换为完整扩展的文本表示，提升TTS输出的自然度。

📦 安装指南

文档中未提及具体安装步骤，可参考使用示例中的依赖安装部分：

poetry add transformers huggingface_hub onnxruntime-gpu torch

📚 详细文档

模型描述

mbart-large-50-verbalization 是 facebook/mbart-large-50 模型的微调版本，专为将乌克兰语文本进行口语化处理而设计，以便用于文本转语音（TTS）系统。该模型旨在将数字和日期等结构化数据转换为乌克兰语的完整扩展文本表示。

架构

该模型基于 facebook/mbart-large-50 架构，该架构在众多语言的翻译和文本生成任务中效果显著。

训练数据

模型在Ubertext数据集中的457,610个句子子集上进行了微调，重点关注新闻内容。口语化的等效文本使用Google Gemini Pro创建，为学习文本转换任务提供了丰富的基础。 数据集 skypro1111/ubertext-2-news-verbalized

训练过程

模型进行了410,000个训练步骤（1个epoch）。

from transformers import MBartForConditionalGeneration, AutoTokenizer, Trainer, TrainingArguments
from datasets import load_dataset, DatasetDict
import torch

model_name = "facebook/mbart-large-50"

dataset = load_dataset("skypro1111/ubertext-2-news-verbalized")
dataset = dataset.train_test_split(test_size=0.1)
datasets = DatasetDict({
    'train': dataset['train'],
    'test': dataset['test']
})

tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.src_lang = "uk_XX"
tokenizer.tgt_lang = "uk_XX"


def preprocess_data(examples):
    model_inputs = tokenizer(examples["inputs"], max_length=1024, truncation=True, padding="max_length")
    with tokenizer.as_target_tokenizer():
        labels = tokenizer(examples["labels"], max_length=1024, truncation=True, padding="max_length")
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

datasets = datasets.map(preprocess_data, batched=True)

model = MBartForConditionalGeneration.from_pretrained(model_name)

training_args = TrainingArguments(
    output_dir=f"./results/{model_name}-verbalization",
    evaluation_strategy="steps",
    eval_steps=5000,
    save_strategy="steps",
    save_steps=1000,
    save_total_limit=40,
    learning_rate=2e-5,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    num_train_epochs=2,
    weight_decay=0.01,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=datasets["train"],
    eval_dataset=datasets["test"],
)

trainer.train()
trainer.save_model(f"./saved_models/{model_name}-verbalization")

性能

该模型未明确使用评估指标，其性能主要通过在提升TTS输出自然度方面的应用来体现。

局限性和伦理考量

用户应注意模型在理解高度细微或特定领域内容时可能存在的局限性。在将该模型部署到实际应用中时，包括公平性和偏差在内的伦理考量也至关重要。

引用

Ubertext 2.0

@inproceedings{chaplynskyi-2023-introducing,
  title = "Introducing {U}ber{T}ext 2.0: A Corpus of Modern {U}krainian at Scale",
  author = "Chaplynskyi, Dmytro",
  booktitle = "Proceedings of the Second Ukrainian Natural Language Processing Workshop",
  month = may,
  year = "2023",
  address = "Dubrovnik, Croatia",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2023.unlp-1.1",
  pages = "1--10",
}

mBart-large-50

@article{tang2020multilingual,
    title={Multilingual Translation with Extensible Multilingual Pretraining and Finetuning},
    author={Yuqing Tang and Chau Tran and Xian Li and Peng-Jen Chen and Naman Goyal and Vishrav Chaudhary and Jiatao Gu and Angela Fan},
    year={2020},
    eprint={2008.00401},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

🔧 技术细节

• 模型基于 facebook/mbart-large-50 架构进行微调，针对乌克兰语文本的口语化处理进行了优化。
• 训练数据来自Ubertext数据集的新闻内容子集，使用Google Gemini Pro生成口语化等效文本。
• 训练过程中进行了410,000个训练步骤（1个epoch），并设置了相应的训练参数，如学习率、批次大小等。

📄 许可证

该模型根据MIT许可证发布，与基础的mbart-large-50模型保持一致。