Downloads · 30 days
13
9% of all-time downloads
Morton-Li/QiDeBERTa-small
QiDeBERTa-small is a machine learning model from Morton-Li. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Model Name: QiDeBERTa-small Language: Chinese Architecture: DeBERTa V3 (modified) Author: Morton Li License: Apache-2.0 Tags: deberta, mlm, chinese, language-modeling, robustness, general-purpose, pretraining, masked-…
Downloads · 30 days
13
9% of all-time downloads
All-time downloads
146
Public
Parameters
38.5M
155 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors154 MB · 100%
From the Hugging Face model README
Model Name: QiDeBERTa-small
Language: Chinese
Architecture: DeBERTa V3 (modified)
Author: Morton Li
License: Apache-2.0
Tags: deberta, mlm, chinese, language-modeling, robustness, general-purpose, pretraining, masked-language-modeling
QiDeBERTa 是一个专为中文构建的通用语言理解预训练模型,采用改进的 DeBERTa V3 架构,强调对语言结构的深度建模能力与对输入扰动的鲁棒性。该模型在多个典型评估基准中表现出色,展现出优异的泛化能力,尤其适用于需要对中文语义细节进行精密捕捉的复杂 NLP 任务。
QiDeBERTa 的设计目标是为中文自然语言处理提供一个稳定、可扩展、具备良好迁移性能的基础模型,可为后续各类下游任务提供强有力的编码支持。
QiDeBERTa 使用 SentencePiece 实现分词器(SPM),训练于与预训练语料相同来源的高质量中文文本数据集之上,对《通用规范汉字表》一级、二级、三级汉字实现 100% 字符覆盖率。
分词器可通过 HuggingFace Tokenizer API 直接加载。
模型预训练语料来源于多个高质量中文语料源,包括但不限于:
所有语料均为公开来源或基于合规机制采集,确保数据合规性、多样性与覆盖性。
QiDeBERTa 使用 Masked Language Modeling 作为核心训练目标,并配合多策略输入扰动增强机制以提高模型泛化能力。通过不同类型的扰动样本搭配与之匹配的任务头进行混合建模,模型具备良好的语义鲁棒性与错误容忍能力。
此外,引入目标导向的训练损失调控机制,强化模型在输入变异与原始输出接近场景下的学习效果,进一步提升其泛用性能。
weight_decay=0.01, LayerNorm & bias权重例外)3e-5,线性衰减,warmup 6%from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("Morton-Li/QiDeBERTa-small", trust_remote_code=True)
model = AutoModel.from_pretrained("Morton-Li/QiDeBERTa-small", trust_remote_code=True)
texts = [
"这是一个测试句子,用于验证语言模型的性能。",
"我爱北京天安门,天安门上太阳升。"
]
outputs = model(**tokenizer(texts, padding=True, return_tensors='pt')) # BaseModelOutput[last_hidden_state, hidden_states, attentions]
If you find this project useful in your research, publications, software development, or any other work, please cite it as follows:
@misc{QiDeBERTa-small,
author = {Morton Li},
title = {QiDeBERTa: Chinese Pretrained DeBERTa Model},
year = {2025},
}
Apache License 2.0