inclusionAIOfficial·12 Agent Skills Included

AReaL

Train large language models with fast asynchronous reinforcement learning

Runs large-scale asynchronous reinforcement learning to train reasoning and agentic LLMs on GPU clusters. Removes the pain of wiring distributed training, rollout workflows, reward functions, and dataset loaders by hand. Ships ready-made skills for adding models, rewards, datasets, tests, and debugging distributed training so agents complete ML engineering tasks faster.

5,708|588

↗

npx skills add inclusionAI/AReaL --all -g -y

Available:

Tells the AI agent how to navigate the AReaL codebase, which expert subagents and skills to invoke for each task, and what coding, testing, and distributed-training rules to follow.

All Skills in This Repository (12)

Pure Emerald Level Indicators

📦 In Repo

inclusionAI

add-reward

Implement and register reward functions in the AReaL reinforcement learning framework.

add-dataset

Scaffold and register AReaL dataset loaders for SFT and RL workflows.

add-workflow

Add a RolloutWorkflow to the AReaL codebase with registration and tests.

add-unit-tests

Create unit tests for AReaL using templates and naming conventions.

debug-distributed

Diagnose distributed training hangs, OOM, and NCCL errors in AReaL.

add-archon-model

Guides adding new HuggingFace model architectures to the Archon training engine.

commit-conventions

Generates Conventional Commits messages with scope inferred from AReaL file paths.

add-unit-tests

Guides writing pytest unit tests for AReaL following project testing conventions.

translate-doc-zh

Translates English Markdown documentation into matching Chinese documents under docs/zh.

upgrade-deps

Upgrade pinned runtime dependencies in AReaL with checklist-driven API compatibility audits.

review-pr

Reviews pull requests with risk classification, domain checklists, and expert subagent consultation.

create-pr

Rebases, squashes, and pushes a branch to create or update a GitHub pull request.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install AReaL?▼

Run `npx skills add inclusionAI/AReaL --all -g -y` in your terminal to install all skills in this suite globally.

What is AReaL used for?▼

AReaL is a distributed reinforcement learning framework for training large language models, especially reasoning and agentic models, with high-throughput asynchronous training.

How do I add a custom reward function in AReaL?▼

Use the add-reward skill, which walks the agent through creating the reward file, registering it, wrapping blocking calls, and adding tests.

Can AReaL train agents built with other frameworks?▼

Yes. Any agent framework like OpenAI Agents SDK or LangChain works by implementing a simple async run method that returns a reward.

Does AReaL support multi-GPU and multi-node training?▼

Yes. It supports FSDP2, Megatron, and Archon engines with local, Ray, and Slurm launchers, plus a debug-distributed skill for troubleshooting hangs and OOM errors.