开发者发布了WangNet，一个用于判断数字是否为Numberwang的开源神经网络模型1。该项目以极轻的体量著称，整个模型仅为1.8 MB的JSON文件，包含约80,804个参数1。推理代码采用纯Python标准库实现，仅需约100行代码，用户可直接克隆运行，无需安装任何外部依赖1。
模型架构采用字符级输入，经过32维嵌入层、两层1D卷积（均为128个核、卷积核大小为3）和全局最大池化后接入两层全连接层，最后通过softmax输出1。在486个测试样本上，模型达到88.9%的准确率，macro-F1得分为0.8961。该模型的主要弱点在于算术运算方面，准确率仅为44–72%1。WangNet支持11种语言，采用MIT许可证发布，可在本地运行或通过Hugging Face Spaces托管部署1。
A developer has released WangNet, a compact neural network model designed to determine whether a number qualifies as Numberwang.1 The entire model is distributed as a 1.8 MB JSON file containing approximately 80,804 parameters, with inference code requiring only about 100 lines of pure Python standard library code.1 The project supports 11 languages and requires zero external dependencies, allowing users to clone and run it directly.1
The model achieves 88.9% accuracy across 486 test samples, with a macro-F1 score of 0.896.1 Its architecture processes character input through an embedding layer (32 dimensions), followed by two convolutional layers with ReLU activation, global max pooling, and a linear classification head with softmax output.1 However, the model shows weakness in arithmetic operations, with accuracy ranging from 44–72% on such tasks.1 The project is licensed under MIT and can be deployed locally or hosted on Hugging Face Spaces.1
评论
还没有评论，欢迎留下第一条。