菜单
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.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/inclusionAI/AReno --skill areno-add-algorithm命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Autonomously research, implement, train, evaluate, visualize, and submit a genuinely new AReno RLVR mini-game whose mechanically generated samples are independent state-to-single-output tasks. Use when the user requests a complete new game demo with real reward improvement, not for multi-turn agent workflows, SFT-only examples, or design-only proposals.
Add or debug an AReno model family, including config conversion, module construction, checkpoint load/save, text or multimodal inference, training backward, tensor parallelism, CUDA graph decode, and checkpoint round trips. Use only when model-specific implementation or compatibility changes are required.
Create or debug an AReno multi-turn agentic dataset, run_agent implementation, tool schemas, tool execution, reward, loss masks, interactive TUI game with OpenAI-compatible LLM inference, or agentic training example. Use for text or multimodal agentic workflows, not ordinary single-turn rollout.
基于 SOC 职业分类
正在显示 SKILL.md
| 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. |
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
areno/api/loss_fns/ or advantage helpers.AlgorithmSpec; load experimental implementations through areno/experimental/ when appropriate.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.