LMOps
A Microsoft research initiative on fundamental technology for building AI products with large language and multimodal models, covering prompt optimization, long context, alignment, faster inference and domain adaptation.
LMOps is a Microsoft research initiative on the general technology needed to build AI products with large language models and multimodal models. It groups its work into themes including better prompts, such as automatic prompt optimization, extensible prompts and in-context example retrieval; longer context, through structured prompting and length-extrapolatable transformers; LLM alignment using feedback; faster inference through lossless acceleration; and adapting a model to a domain.
Related Microsoft projects linked from the repository include unilm, for self-supervised pretraining across tasks, languages and modalities, and torchscale, for scaling Transformer architectures. The repository mainly links research papers, with some accompanying code and model releases.