Snowflake previews Cortex Training with OpenHands Snowflake previews Cortex Training with OpenHands

The training loop stays configurable

Cortex Training supports supervised fine-tuning and reinforcement learning, with either full-parameter updates or LoRA adapters. Snowflake names TRL, veRL and SkyRL among the supported frameworks. Developers can also write custom loops using APIs for generation, forward and backward passes, optimizer steps and weight synchronization. The service manages the underlying GPUs and checkpoints, while the training script retains control over learning. Its open-source backend is Arctic Platform. Trained checkpoints can be loaded for evaluation or deployed through OpenAI-compatible endpoints backed by dedicated GPUs. Snowflake bills by GPU time. Server-side checkpoints persist so teams can resume training or launch further experiments from earlier states.

Train inside the configured harness

OpenHands describes rollouts that retain the agent’s prompts, tools, skills and execution limits. A proxy records the exact tokens sampled across chat-completion turns. Assistant outputs contribute to the training loss; prompts and tool results do not. Harbor tasks supply instructions, an environment and a test script whose result becomes the reward. In the companies’ demonstration, Qwen3.5-4B located relevant code in Python repositories over twenty-five steps, with sixty-four rollouts per step. The score combined file, module and function F1. This measures code localization within the configured harness. Compatibility has limits: the file editor is unavailable during training. Terminal commands use fresh shells, retaining the working directory but losing exported variables and background jobs.

Evaluation is an explicit setting

The supporting Harbor recipe, revised November 1, documents operating conditions that affect how a run should be read. Before allocating GPUs, the launcher exercises the first task with a do-nothing agent, checking sandbox startup, environment preparation and grading. Evaluation during training is disabled unless requested. Its evaluation-task path defaults to the training tasks, so a separate held-out directory must be supplied for an independent test set. A rollout that ends without a verifier result is recorded as an error and excluded from the loss, rather than assigned a zero reward. If every rollout in a step fails, the recipe retries up to three times before stopping. The default run also makes one pass over the task set. A larger requested step count does not override that dataset limit; additional epochs require a separate setting. Installed Harbor agents are refused because they would expose the model endpoint key inside task environments; the recipe instead requires a driver-side agent.