qwen_ner

Qwen2 -> Named Entity Recognition

Note

With the latest update, the pipeline now supports high-level training control through a YAML file, eliminating the need to modify source code, except when adding a new dataset. For new datasets, you must still convert your dataset to the required format. For all other cases, simply modify the configuration file (i.e., train_config.yaml) to make necessary adjustments.

Basic Information

Setup Details

• Accelerator: NVIDIA RTX 4090D $\times$ 2
• Platform: Linux
• Internet: Enabled

Model and Resources

• LLM: Qwen/Qwen2-1.5B-Instruct
• Dataset: The Learning Agency Lab - PII Data Detection
• Utils: transformers | trl

Additional Information

Important

Key modules are implemented in the qwen2ner module. For more technical details, please refer to the module.

• Download the dataset from Kaggle to the dataset folder.

• Construct the .csv format dataset.

python3 construct_text_data.py

• Train the model.

./train.sh

• Inference on a single text.

python3 inference.py \
    --model_name_or_path MODEL_NAME_OR_PATH

Blogs / 中文博客

• 知乎：【大模型微调】Qwen SFT：基于 trl 框架的 QLoRA 微调
• CSDN：【大模型微调】Qwen SFT：基于 trl 框架的 QLoRA 微调