| Core Concepts | The four objects in the training loop and the four-knob invariant. |
| Argument Groups | Where MODEL_ARGS, PERF_ARGS, GRPO_ARGS, and the other launch-script arrays belong. |
| Training Backend | Megatron-LM as the training backend — parallelism, checkpoints, and hooks. |
| Training Script Walkthrough | The eight XXX_ARGS arrays in a launch script, plus the execution modes (sync/async, colocation, dynamic sampling, partial rollout, BF16+FP8). |
| Monitoring & Logging | wandb, structured logs, per-source breakdowns, profiling, router metrics. |
| Customization | The 21 --*-path plug-points for custom Python — rollout, reward, filters, loss, hooks. |
| Rollout Endpoints | The /generate endpoint and the OpenAI chat endpoint for agentic sessions. |
| Fully Async Rollout | Queue-backed rollout production, tuning knobs, and when to use train_async.py. |
| Agentic Chat Templates | Turning on and verifying TITO so multi-turn agentic rollout stays append-only. |
| CLI Reference | Every flag Miles accepts, grouped by subsystem. |
| Environments | Supplying an environment: dataset + reward, your own env via the plug points, or an external ecosystem. |
