GangLLM is a large language model project specifically designed for arguing, and the model is named "GangDan'er." It can produce contrarian-style replies based on user input, is capable of catching subtle loopholes in users' wording, and launches sharp rebuttals. The model is fine-tuned from an open-source large model; the base model is InternLLM-Chat-7B, and it is fine-tuned and trained using 1,680 contrarian-style dialogue data entries plus 100 self-cognition data entries. The project specifically warns: those with heart problems or mental illness should use it with caution, as the author is not responsible for psychological or physiological problems caused by using this model.
Main capabilities: The model focuses on generating argumentative replies, can identify loopholes in users' speech, and rebut them.
Usage: Two methods are provided. The first is to run python app/app.py; the second is to use xtuner chat ./merged --prompt-template internlm_chat. During training, use the xtuner command: xtuner train internlm_chat_7b_qlora_self.py --deepspeed deepspeed_zero2.
Applicable scenarios: Suitable for scenarios requiring an argumentative dialogue experience or research on generating contrarian replies.
Limitations: Hardware requirements are an A100 40G GPU, 100G hard drive, and more than 32G memory; the software environment requires Anaconda and CUDA/CUDNN. The base model is InternLM2_7B_chat. The project code and model files use the Mulan PSL, allowing free modification and commercial use. Dataset sources include the Ruozhiba post dataset and others; specific details are not yet provided. For more information, you can visit the Bilibili video, OpenXLab, and ModelScope links provided by the project.
