The focus of tool stacks in the era of large models is different from that in the era of traditional machine learning
This chapter starts from the workflow perspective of enterprise training models and building AI-enabled applications, explains the main links involved in detail, and pays attention to the differences in the process of LLMOps and MLOps. We believe that AI = Data + Code, after data preparation, model training, model deployment, product integration, look at the following steps:
► Data preparation: high-quality labeled data, continuous demand for feature databases, and synthetic data may become a future trend. Data preparation is a time-consuming and important part in both traditional MLOps and LLMOps. Unsupervised learning reduces the demand for labeled data, but the RLHF mechanism embodies the importance of high-quality labeled data. We believe that the demand for massive training data for ultra-large parameter models in the future may be met by synthetic data. In addition, Data+AI platform manufacturers are key.
► Model training: The model library is more rigid, the training framework continues to iterate, and software tools assist in experiment management. Fine-tuning and distilling small models based on the general LLM large model has become a cost-effective landing method. Therefore, a model library that can efficiently and conveniently obtain pre-trained models is required; it also leads to the underlying distributed computing engine and training that are more suitable for the large-scale training needs of LLM. frame. Additionally, we believe experiment management tools are likely to remain high in importance.
