verl-rl-training

Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.

Quality

82%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Quality

Content

65%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is highly actionable with executable code and clearly sequenced workflows, but it is padded with marketing-style framing and underuses its own bundle files — including a dangling reference to a missing file. Tightening the overview sections and wiring up the existing references would materially improve it.

Suggestions

Remove or condense the "When to Use verl"/"Consider alternatives" and "Key Features" sections so the body leads with action rather than restating the description.

Replace the inline "Common Issues and Solutions" block with a signaled link to references/troubleshooting.md, and add a link to references/api-reference.md from the Configuration Reference section so the existing bundles are actually used.

Fix the dangling reference to references/multi-turn.md — either create the file or remove the broken link.

Dimension	Reasoning	Score
Conciseness	Mostly efficient with real configs, commands, and code, but the "When to Use"/"Consider alternatives" and "Key Features" marketing-style sections restate the description and could be trimmed without losing actionability.	2 / 3
Actionability	Provides fully executable, copy-paste-ready guidance throughout: `python3 -m verl.trainer.main_ppo` invocations, complete YAML configs, a working reward function, and multi-node `ray start` commands.	3 / 3
Workflow Clarity	Workflows 1–3 are clearly numbered with prerequisite checklists and a Step 5 validation checkpoint, but validation is soft (no explicit validate→fix→retry loop) for operations that should confirm reward/loss health before proceeding.	2 / 3
Progressive Disclosure	Bundle files exist (api-reference.md, troubleshooting.md) but are never referenced from the body, while the inline "Common Issues" and "Configuration Reference" duplicate what belongs in them; additionally the body links to references/multi-turn.md which does not exist.	2 / 3
	Total	9 / 12 Passed

Description

100%

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A concise, third-person description that clearly states both capabilities and explicit trigger conditions with strong natural-language keywords. It is among the better examples and requires no significant changes.

Dimension	Reasoning	Score
Specificity	Names multiple concrete actions and domains — "training LLMs with reinforcement learning", "RLHF, GRPO, PPO", "post-training at scale" — matching the anchor for listing several specific concrete actions.	3 / 3
Completeness	Explicitly answers both "what" ("Provides guidance for training LLMs with reinforcement learning using verl") and "when" ("Use when implementing RLHF, GRPO, PPO, or other RL algorithms"), with an explicit trigger clause.	3 / 3
Trigger Term Quality	Includes the natural terms a practitioner would actually say ("RLHF, GRPO, PPO", "RL algorithms", "post-training", "at scale"), giving good coverage of likely user phrasing.	3 / 3
Distinctiveness Conflict Risk	Has a clear niche (verl / Volcano Engine RL training for LLMs) with distinct triggers unlikely to fire for unrelated skills.	3 / 3
	Total	12 / 12 Passed

Validation

81%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
relative_links	Relative link issues: 1 missing	Warning
referenced_paths_exist	Referenced path issues: 2 missing	Warning

	Total	13 / 16 Passed

Repository: Orchestra-Research/AI-Research-SKILLs
Path: 06-post-training/verl/SKILL.md
Commit: 773a529

Reviewed: about 6 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.