AWS Wordle guide separates training scores from serving results

The AWS walkthrough uses Wordle as a teaching example and distinguishes training validation from evaluation at serving settings. Changing reasoning effort for evaluation requires editing the agent’s environment variables and redeploying it. That instruction has an important boundary: the authors warn against redeploying during a running training job, because restarting the runtime interrupts ongoing attempts and can fail the job.

The method guide defines multi-turn reinforcement learning as training decisions across a sequence. An agent acts, receives feedback and continues from the changed situation; the goal is cumulative reward rather than an isolated answer. An episode means the complete task, while a turn pairs an action with the environment’s response. This approach is relevant when later tool calls depend on earlier results, or individual steps work but the complete task does not.

For comparison, the evaluation documentation recommends testing the base and trained models on prompts excluded from training. Keep that evaluation set stable across iterations, preserve the prompt format, and cover tool combinations and previous failures. Inspect the recorded action sequences and reward distributions in MLflow, alongside turn counts. Many turns can indicate looping or excessive retries; the count alone does not establish the cause. Raw evaluation artifacts are available in the configured S3 output location.