- Online optimization entrypoints (
optimize_teaching.py, runtime adapters, optimize shell scripts, and configs/optimize assets). - Online optimization notebooks and Mintlify pages that referenced the deprecated workflow.
scripts/run_offline_pipeline.pyhelper for one-touch export → GRPO training workflows.- GRPO-focused example configs (
configs/examples/quickstart.yaml,configs/demo/runtime_grpo.yaml) aligned with runtime trace exports. - Conditional dependency handling for
bitsandbytes(Linux-only, matching upstream wheels) and explicitlitellmrequirement so the reward tests install cleanly on CPU dev machines.
- Repositioned Atlas Core documentation and README to describe the repo as offline GRPO + reward tooling, directing online continual learning to the atlas-sdk runtime.
- Updated Hydra defaults and configs to use runtime trace datasets and GRPO training out of the box.
- Refreshed docs navigation and quickstarts to highlight the export → train → deploy pipeline.