Applied Compute Agent Cloud (AC2) has launched an end-to-end infrastructure designed to streamline the process of training and inferencing open weight models. This comprehensive platform serves as a researcher cockpit, facilitating model development, inference, and continual learning. By providing a unified environment, AC2 aims to simplify the complexities associated with model training and deployment.
To get started with AC2, users can leverage a coding agent such as Claude Code, Cursor, or Codex. By pasting a specific prompt into one of these agents, the setup process is automated, installing the CLI, logging in, and configuring the AC2 MCP server. Alternatively, users can follow the Quickstart guide to manually run through the steps.
AC2 comprises several key components, including the ac2.sdk, ac2.runtime, and ac2.tracing. The ac2.sdk functions as a platform client, managing datasets, workloads, sessions, traces, and secrets. The ac2.runtime is an agent framework that encompasses agents, environments, orchestrators, tasks, graders, and episodes. Furthermore, ac2.tracing utilizes OpenTelemetry for observability, allowing users to capture timing, inputs, outputs, and nested spans by adding the @traced decorator to any function.
With AC2, users can install the SDK and train a model, define agents and environments, and manage datasets and deployments from Python. The platform also enables the analysis of evaluation and training traces directly from the editor, facilitating the running of evaluations, grading of traces, and comparison of behavior. Users can choose from various training methods, including GRPO, supervised fine-tuning, and on-policy self-distillation. Additionally, AC2 allows for model deployment and traffic routing through policies and endpoints.
In terms of troubleshooting, AC2 provides resources to help users find answers for failed runs, dataset issues, package installation problems, and cleanup. The platform's documentation and support materials are generated using AI, which may occasionally contain mistakes.
Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.