| name | areno-add-algorithm |
| description | Add or modify an AReno algorithm, trainer, loss, advantage calculation, role model, or algorithm-specific configuration. Use for framework-level SFT, DPO, GSPO, GRPO, PPO, or new optimization method development. Do not use merely to run an existing algorithm. |
Add an AReno Algorithm
Start from AlgorithmSpec and registration in areno/api/algorithms.py; do not add factory branches.
Develop on a dedicated local branch. Commit changes locally, then update the
remote GPU checkout by fetching and pulling that branch; never patch source on
the remote host. Use ModelScope for any model or dataset references used by the
validation workload.
python .agents/skills/areno-add-algorithm/scripts/inspect_algorithms.py
Workflow
- Define whether the algorithm is offline, rollout policy-only, or multi-role. Read references/ownership.md.
- Specify input records, sequence construction, masks, role models, loss inputs, and metrics before coding.
- Add the narrowest config type and preserve public defaults/compatibility.
- Put batch/materialization logic in the trainer and tensor mathematics in
areno/api/loss_fns/ or advantage helpers.
- Register one
AlgorithmSpec; load experimental implementations through areno/experimental/ when appropriate.
- Add CPU tests for registration, config, masks, exact small-tensor math, and trainer dispatch.
- Run the new algorithm end to end for at least two consecutive successful training steps using a real model and representative data. Verify finite losses, metrics, and gradients on both steps. For rollout algorithms, also verify bounded GPU rollout/train logprob consistency.
Completion requires registry discovery, deterministic mathematical tests, evidence from at least two successful end-to-end training steps, and role lifecycle checks where applicable. A one-step smoke train is useful for diagnosis but does not complete algorithm validation.
