Fengshenbang-LM

Why we included this project

Chinese-language NLP teams often find that the best open models skew English-first, which is exactly the gap this umbrella release from IDEA Research tries to close. Fengshenbang is not one model but a family of pretrained checkpoints with Chinese strength, from the Ziya general-purpose chat models to the Erlangshen encoders for understanding tasks and generation-focused variants like Wenzhong and Randeng. The repo also carries the Fengshen training and fine-tuning framework with example scripts, so you can take a released checkpoint and adapt it to your own downstream task instead of starting from scratch. Since weights are mirrored on Hugging Face, plugging a pretrained checkpoint into an existing pipeline takes only a few lines of code. For teams building Chinese chatbots, text classification, extraction, or multimodal features, it is a practical set of baselines to benchmark against rather than a single toy demo.