标签:

• 由训练器自动生成 数据集: UrukHan/wav2vec2-russian 小部件示例:
• 文本: "西方在俄罗斯启动乌克兰非军事化特别行动后，实施了多轮新经济制裁。克里姆林宫称新限制措施严厉，但表示俄罗斯已提前做好准备。" 模型索引:
• 名称: t5-russian-summarization 结果: []

t5-russian-summarization

用于校正音频转写文本的模型。我的音频识别模型位于https://huggingface.co/UrukHan/wav2vec2-russian，其输出结果可输入本模型处理。测试数据来自随机YouTube视频。

训练数据集: UrukHan/t5-russian-summarization：https://huggingface.co/datasets/UrukHan/t5-russian-summarization

使用示例及Colab操作指南

完整带注释的Colab笔记本：https://colab.research.google.com/drive/1ame2va9_NflYqy4RZ07HYmQ0moJYy7w2?usp=sharing

# 安装transformers库
!pip install transformers

# 导入库
from transformers import AutoModelForSeq2SeqLM, T5TokenizerFast

# 设置模型名称
MODEL_NAME = 'UrukHan/t5-russian-summarization'
MAX_INPUT = 256

# 加载模型和分词器
tokenizer = T5TokenizerFast.from_pretrained(MODEL_NAME)
model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)

# 输入数据（支持多文本列表或单文本）
input_sequences = ['西方在俄罗斯启动乌克兰非军事化特别行动后...'] 

task_prefix = "拼写校正: "
encoded = tokenizer(
  [task_prefix + seq for seq in input_sequences],
  padding="longest",
  max_length=MAX_INPUT,
  truncation=True,
  return_tensors="pt",
)

predicts = model.generate(encoded)    # 生成预测
tokenizer.batch_decode(predicts, skip_special_tokens=True)  # 解码结果

训练配置指南

定制化Colab训练笔记本：https://colab.research.google.com/drive/1H4IoasDqa2TEjGivVDp-4Pdpm0oxrCWd?usp=sharing

# 环境配置
!pip install datasets transformers sentencepiece rouge_score
!apt install git-lfs

# 核心训练逻辑
from transformers import Seq2SeqTrainer, T5TokenizerFast, AutoModelForSeq2SeqLM

# 参数设置
REPO = "t5-russian-summarization"
MODEL_NAME = "UrukHan/t5-russian-summarization" 
MAX_INPUT = 256
MAX_OUTPUT = 64

# 数据预处理示例
def preprocess_function(examples):
    inputs = [prefix + doc for doc in examples["document"]]
    model_inputs = tokenizer(inputs, max_length=MAX_INPUT, truncation=True)
    with tokenizer.as_target_tokenizer():
        labels = tokenizer(examples["summary"], max_length=MAX_OUTPUT, truncation=True)
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

数据格式转换示例

# 样本数据
input_data = ['西方在俄罗斯...']
output_data = ['西方实施新制裁']

# 输入编码
task_prefix = "拼写校正: "
encoding = tokenizer(
    [task_prefix + seq for seq in input_data],
    padding="longest",
    max_length=MAX_INPUT,
    truncation=True,
    return_tensors="pt"
)

# 标签处理
labels = tokenizer(output_data, padding="longest", max_length=MAX_OUTPUT, truncation=True).input_ids
labels = torch.tensor(labels)
labels[labels == tokenizer.pad_token_id] = -100  # 忽略填充符