​Post-training is becoming a critical workload for data providers. As data and environment supply increases, it is increasingly important to validate and show the quality of the data, which means vendors are training their own models: orchestrating evals, rollouts, sandboxes, and training and inference workloads all at once.

We're excited to gather practitioners in the community together to share best practices and hear how Mercor is running RL at scale with SkyRL; how Harbor is providing the open standard for evaluation and optimization; and how SkyRL is pushing the frontier on enabling large scale post-training for all

​Agenda:

​5:30 PM – Doors open, food, and networking

​6:00 PM – SkyRL at Mercor (Tianhua Tao ,MTS at Mercor, Research Team)

​6:20 PM – What’s new in SkyRL: 1M-context training, TITO proxy, Ray Sandboxes & more (Avi Basnet, Software Engineer, Anyscale, Charlie Ruan, SkyRL Core contributor and Xinyu Zhang, Software Engineer, Anyscale)

​6:40 PM – Software Engineering Became Machine Learning (Alex Shaw, co-creator Harbor and Terminal-Bench)

​7:00 PM – Open networking + happy hour

8:30 PM – Event Ends

​Hosted by:

​Anyscale, the company behind the Ray open source project, is a fully managed AI platform for foundation model-scale workloads. AI teams use it to run data, training, inference, and RL workloads at scale on any cloud.

​Harbor builds open-source infrastructure for specifying sandboxed agent tasks for evaluation and optimization, from the makers of Terminal-Bench.

​UC Berkeley Sky Computing Lab is the latest in Berkeley EECS’s series of five-year collaborative research labs, working on AI and systems.

​Mercor is the AI data company for building AI systems with human intelligence. Its network of more than five million experts helps frontier AI labs, big tech, AI startups, and Fortune 500 companies train, evaluate, and measure AI on real professional work. Founded in 2023 and headquartered in San Francisco, Mercor is backed by Benchmark, Felicis, General Catalyst, and Menlo Ventures.