slime

RL post-training for LLMs with Megatron and SGLang.

Quality

53%

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 ./optional-skills/mlops/slime/SKILL.md

The canonical home for this skill is slime-rl-training in Orchestra-Research/AI-Research-SKILLs

Quality

Content

57%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 body is highly actionable and well-structured for a complex RL training framework, with concrete commands and clear workflow steps. Its main weaknesses are duplicated content (Quick Start vs Workflow 1, inline issues vs the troubleshooting reference) and orphaned bundle reference files that are never signaled from the body.

Suggestions

Replace the inline 'Common Issues and Solutions' section with a pointer to references/troubleshooting.md, and move the bulk of 'Configuration Reference' to references/api-reference.md to use the existing bundle files.

Collapse the 'Quick Start: GRPO Training' and 'Workflow 1: Step 3' duplicate train.py blocks into one canonical example referenced from both places.

Add an explicit validation feedback loop in Workflow 1 (e.g. check reward curves / loss before continuing past the first N rollouts, and what to do if they diverge) to satisfy the batch-operation validation requirement.

Dimension	Reasoning	Score
Conciseness	The body is dense with genuinely slime-specific content and avoids explaining concepts Claude already knows, but 'Quick Start: GRPO Training' and 'Workflow 1: Step 3' duplicate near-identical train.py invocations, and the inline 'Common Issues' section duplicates references/troubleshooting.md.	3 / 5
Actionability	It provides copy-paste-ready docker, git, source, and train.py commands plus concrete data-format examples throughout, with only minor gaps such as the undefined helper functions in custom_generate.py and reliance on externally-sourced ${MODEL_ARGS[@]}.	4 / 5
Workflow Clarity	Workflows 1-3 are clearly sequenced with prerequisite and monitoring checklists, but this long-running batch training workflow lacks explicit validate->fix->retry feedback loops within the steps, capping workflow clarity at 3 for batch operations.	3 / 5
Progressive Disclosure	Section structure is clear and well-organized, but the two bundle reference files (api-reference.md, troubleshooting.md) are never linked from the body and their content is inlined instead, leaving the references orphaned and content that belongs in separate files inline.	3 / 5
	Total	13 / 20 Passed

Description

48%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 is specific and distinctive for a niche technical audience but is purely a noun phrase: it states what slime is without listing concrete actions or any 'use when' trigger guidance. Adding a verb-led capability list and a trigger clause would lift completeness and specificity.

Suggestions

Lead with concrete actions, e.g. 'Trains LLMs with RL post-training (GRPO/PPO) using Megatron-LM and SGLang rollouts'.

Append a 'Use when...' clause naming natural trigger phrases such as reinforcement learning, RLHF, GRPO, reward-model fine-tuning, or Megatron/SGLang setup.

Add common synonyms (reinforcement learning, RLHF, fine-tuning) so users searching with those terms match the skill.

Dimension	Reasoning	Score
Specificity	The phrase 'RL post-training for LLMs with Megatron and SGLang' clearly names the domain and supporting tools but contains no concrete action verbs (e.g. train, generate rollouts, fine-tune), so it sits at 'names domain but actions are minimal'.	2 / 5
Completeness	It gives a clear 'what' (RL post-training for LLMs via Megatron + SGLang) but provides no 'Use when...' trigger guidance, which caps completeness at 3 per the rubric guidelines.	3 / 5
Trigger Term Quality	It surfaces core technical keywords (RL, post-training, LLMs, Megatron, SGLang) that specialists would say, but omits common synonyms and variations such as reinforcement learning, RLHF, GRPO, or fine-tuning.	3 / 5
Distinctiveness Conflict Risk	Naming two specific frameworks alongside RL post-training carves a distinct niche with minimal conflict risk, but the brevity and missing trigger phrasing keep it just short of a fully distinct, self-triggering description.	4 / 5
	Total	12 / 20 Passed

Validation

75%

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 — 12 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning
metadata_field	'metadata' should map string keys to string values	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
referenced_paths_exist	Referenced path issues: 4 missing, 2 deeper-than-1-level	Warning

	Total	12 / 16 Passed

Repository: NousResearch/hermes-agent
Path: optional-skills/mlops/slime/SKILL.md
Commit: 6680afb

Reviewed: about 4 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.