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

63%

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

Fix and improve this skill with Tessl

tessl review fix ./skills/post-training/verl/SKILL.md

Quality

Content

63%Weight 40%Scale 1-5

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

The content is action-rich with well-sequenced workflows and mostly executable examples, but it carries some redundant verbosity and a dangling reference plus two unreferenced bundle files that hurt navigation.

Suggestions

Fix the broken link to references/multi-turn.md (the file is absent) and add explicit links to the existing references/api-reference.md and references/troubleshooting.md so the bundle is navigable.

De-duplicate the algorithm list that appears in both the 'Key Features' bullets and the 'Algorithm Selection' table to tighten the token budget.

Make validation feedback loops explicit in the training workflows (e.g., 'if reward collapses, reduce lr / raise kl_loss_coef and re-run') rather than only listing post-hoc monitor checkboxes.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient with copy-paste configs and commands, but includes some padding such as repeated algorithm lists across 'Key Features' and the configuration table and the 'Consider alternatives when' editorial section that could be trimmed.	3 / 5
Actionability	It provides executable bash/python/yaml blocks (main_ppo launch commands, dataset prep, reward function, full configs) covering common cases, with only minor gaps such as the multi-node example lacking a concrete head IP and the reward function being illustrative.	4 / 5
Workflow Clarity	Workflows are sequenced with numbered steps and prerequisite/monitoring checklists (e.g., Workflow 1 Steps 1-5 with validation checkboxes), but validation feedback loops (validate -> fix -> retry) are implicit rather than explicit, leaving minor checkpoint gaps.	4 / 5
Progressive Disclosure	The body signals references but links to a non-existent file (references/multi-turn.md) while the actual bundle files (api-reference.md, troubleshooting.md) are never linked from the body, so navigation is incomplete and partly broken.	3 / 5
	Total	14 / 20 Passed

Description

62%Weight 40%Scale 1-5

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

The description clearly states purpose and provides an explicit 'Use when' trigger tied to recognizable RL algorithm names. It is solid but stops short of comprehensive action coverage and natural-synonym breadth.

Suggestions

Add concrete action verbs (e.g., 'orchestrates distributed RL training, runs multi-turn rollouts, swaps backends') instead of the generic 'Provides guidance for training'.

Broaden trigger terms with natural phrasings users actually say, e.g., 'RL fine-tuning', 'policy optimization', 'reward model training', 'RLHF/RLAIF post-training'.

Make the 'when' clause more specific by naming the scale/infrastructure context (multi-GPU/multi-node clusters) so it triggers only in the right scenarios.

Dimension	Reasoning	Score
Specificity	It names the domain ('training LLMs with reinforcement learning using verl') and a couple of concrete actions, but the capability list is generic ('guidance for training') rather than enumerating the concrete operations performed, so it is not comprehensive.	3 / 5
Completeness	It answers both 'what' (guidance for training LLMs with RL via verl) and 'when' ('Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale'), but the 'when' clause could be more explicit and concrete to reach a 5.	4 / 5
Trigger Term Quality	It surfaces real algorithm keywords (RLHF, GRPO, PPO) and 'post-training at scale', but is missing common natural synonyms and phrasings a user might say such as 'RL fine-tuning' or 'policy optimization'.	3 / 5
Distinctiveness Conflict Risk	The verl/RL post-training niche is fairly distinct with specific algorithm triggers, with only minor overlap risk against general distributed-training or SFT/DPO skills.	4 / 5
	Total	14 / 20 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: OpenLAIR/dr-claw
Path: skills/post-training/verl/SKILL.md
Commit: 4b35756

Reviewed: 1 day 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.