BayLing-13B-diff: Cross-lingual Instruction Following LLM

BayLing-13B-diff is a 13 billion parameter instruction-following large language model developed by ICT/CAS, focusing on bridging cross-lingual alignment and instruction following. This model is a weight-diff version of BayLing-13B-v1.0, meaning it provides the differences from a base model, allowing for efficient deployment. It is designed to run on consumer-grade GPUs with 16GB of memory.

Key Capabilities

• Advanced Language Alignment: Demonstrates superior capability in English and Chinese generation.
• Instruction Following: Excels in understanding and executing instructions.
• Multi-turn Interaction: Capable of engaging in extended conversations.
• Resource-Efficient Deployment: Can be deployed on consumer-grade GPUs.
• Multilingual Evaluation: The project includes a human-annotated evaluation set, BayLing-80 Test Set, for assessing multilingual and multi-turn interaction capabilities.

Limitations

BayLing-13B-diff, like other LLMs, has limitations. It may generate inaccurate factual information and lacks proficiency in reasoning, mathematics, and coding tasks. There is also a risk of generating harmful or biased content. For enhanced Chinese knowledge, users are directed to BayLing-13B-v1.1.