supply-chain-bert

基于 bert-base-chinese 微调的中文供应链关系分类模型(PaddlePaddle)。

给定两个行业/产品实体,判断其供应链关系:上游 / 下游 / 其他

模型信息

属性
基础模型 bert-base-chinese
框架 PaddlePaddle
分类头 Linear(768 → 3) + Dropout(0.1)
训练集规模 ~513,000 条三元组
最大序列长度 128

输入格式

[CLS]{entity1}[SEP]{entity2}[SEP]

快速使用

pip install paddlepaddle paddlenlp
import paddle
from paddlenlp.transformers import BertModel, BertTokenizer
from huggingface_hub import hf_hub_download

LABEL_MAP = {"下游": 0, "其他": 1, "上游": 2}
ID_TO_LABEL = {v: k for k, v in LABEL_MAP.items()}

# 下载模型权重
model_path = hf_hub_download("wuhongfei/supply-chain-bert", "best_model.pdparams")

# 加载模型
tokenizer = BertTokenizer.from_pretrained("bert-base-chinese")
bert = BertModel.from_pretrained("bert-base-chinese")

# 推理
entity1, entity2 = "钢铁冶炼", "汽车制造"
text = f"[CLS]{entity1}[SEP]{entity2}[SEP]"
encoded = tokenizer(text, max_length=128, padding="max_length", truncation=True)
input_ids = paddle.to_tensor([encoded["input_ids"]], dtype="int64")
token_type_ids = paddle.to_tensor([encoded["token_type_ids"]], dtype="int64")

或直接使用仓库推理脚本:

git clone https://github.com/wuhongfei/supply-chain-bert
cd supply-chain-bert
python infer.py --model best_model.pdparams --entity1 钢铁冶炼 --entity2 汽车制造

训练配置

超参数
学习率 3e-5
Batch Size 32
Epochs 20
Weight Decay 0.01
Warmup Proportion 0.1
优化器 AdamW

关联资源

License

MIT

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for flipped121364/supply-chain-bert

Finetuned
(282)
this model