Fix: Monkey-patch aistudio_sdk.hub.download before importing PaddleNLP

Fix: Build BertTokenizer from vocab dict to avoid aistudio_sdk

Fix: Load BertTokenizer directly from vocab.txt path

Fix: Build custom BERT classification model from scratch

Fix: Create BertForSequenceClassification directly without from_pretrained

Fix: Load PaddleNLP model state directly to avoid aistudio_sdk

Fix: Use huggingface_hub to download PaddleNLP models

Fix: Use transformers instead of paddlenlp for HuggingFace Hub models

Initial commit - HR Evaluation API v2