slime-rl-training

Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. Use when training GLM models, implementing custom data generation workflows, or needing tight Megatron-LM integration for RL scaling.

Quality

77%

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 ./06-post-training/slime/SKILL.md

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 concrete executable commands and code, but it is over-long because it duplicates material already in the references/ bundle and fails to link to those files, and its training workflows lack explicit validation feedback loops.

Suggestions

Replace the inline 'Configuration Reference', 'Data Buffer System', and 'Common Issues and Solutions' sections with one-line pointers to references/api-reference.md and references/troubleshooting.md, keeping only a concise quick-start in SKILL.md.

Add explicit validation checkpoints to the training workflows (e.g. verify the checkpoint loads, confirm reward curves are non-NaN before continuing, and a 'if OOM/crash → see troubleshooting.md → retry' feedback loop).

Fix dangling references: either add the referenced examples/ and scripts/models/ bundles or change the pointers to files that actually exist in the bundle.

Dimension	Reasoning	Score
Conciseness	The 464-line body is mostly useful and avoids explaining basics Claude already knows, but it reproduces substantial material (Architecture Overview, Configuration Reference, Data Buffer System, Common Issues, Supported Models) that is duplicated and expanded in the references/ bundle, so it could be tightened considerably; not the lean anchor-3.	2 / 3
Actionability	It provides fully executable, copy-paste-ready guidance — complete train.py / train_async.py commands with real flags, python custom_generate and RolloutDataSource examples, and concrete fix commands — matching the anchor-3 'fully executable code/commands' example.	3 / 3
Workflow Clarity	Workflows 1–3 are sequenced with prerequisites and step numbering and a light monitor checklist, but the multi-day training batch operation lacks explicit validation checkpoints and a validate→fix→retry feedback loop; per the rubric, missing validation in batch operations caps this at 2.	2 / 3
Progressive Disclosure	The body is well-sectioned but keeps API-reference and troubleshooting content inline that already exists in references/api-reference.md and references/troubleshooting.md, never signals those actual bundle files, and instead points to examples/ and scripts/models/ paths that are not in the bundle — matching anchor-2 'content that should be separate is inline'.	2 / 3
	Total	9 / 12 Passed

Description

90%

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 strong: it states both capability and explicit use-when triggers in third person, and occupies a clear, low-conflict niche. The only weakness is the generic lead verb 'Provides guidance' instead of naming concrete actions.

Suggestions

Replace the vague verb 'Provides guidance for' with concrete actions, e.g. 'Trains and post-trains LLMs with RL, configures GRPO/PPO rollouts, and implements custom data-generation workflows using slime'.

Consider adding the headline algorithm name (e.g. 'GRPO/RL post-training') to the trigger clause so users searching by algorithm land on this skill.

Dimension	Reasoning	Score
Specificity	It names the domain ('LLM post-training with RL using slime') and several actions ('training GLM models', 'implementing custom data generation workflows', 'Megatron-LM integration'), but the lead verb 'Provides guidance' is vague rather than a list of concrete operations, so it falls short of the multi-action anchor 3.	2 / 3
Completeness	It explicitly answers both what ('Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework') and when ('Use when training GLM models, implementing custom data generation workflows, or needing tight Megatron-LM integration for RL scaling'), matching the anchor-3 example with an explicit 'Use when' clause.	3 / 3
Trigger Term Quality	It covers natural terms a practitioner of this framework would actually say — 'training GLM models', 'custom data generation workflows', 'Megatron-LM integration', 'RL scaling' — giving good coverage of the relevant vocabulary rather than generic jargon.	3 / 3
Distinctiveness Conflict Risk	The niche is highly specific — slime, Megatron+SGLang, GLM models, RL scaling — with distinct triggers that would not realistically fire for unrelated skills, matching the clear-niche anchor 3.	3 / 3
	Total	11 / 12 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
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	14 / 16 Passed

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

Reviewed: about 13 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.