WizardLM-13B is a 13-billion parameter large language model developed by researchers at Microsoft. It is a fine-tuned version of the Llama foundation model series, specifically optimized for complex instruction-following tasks. The model is part of the WizardLM project, which introduced the Evol-Instruction methodology to improve the reasoning and interaction capabilities of open-source language models.
The core innovation behind WizardLM-13B is the use of an LLM to automatically rewrite and evolve simple instructions into more complex, multi-step prompts. This process generates a diverse and difficult dataset that allows the model to better handle intricate human requests. The model has seen multiple iterations, with version V1.2 being a significant update built on the Llama 2 architecture, providing enhanced performance and a larger context window compared to its predecessors.
Evaluation on benchmarks like AlpacaEval and MT-Bench has shown that WizardLM-13B often performs competitively with much larger models. It is designed to follow a Vicuna-style multi-turn conversation format, making it suitable for chat-based applications, creative writing, and summarization tasks that require high-quality instruction adherence.