torchforge-rl-training

Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.

Quality

70%

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/torchforge/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.

A highly actionable skill body with excellent executable examples, held back by missing validation feedback loops in its workflows and by inline API/troubleshooting content that duplicates and ignores the provided reference bundle.

Suggestions

Add explicit validation checkpoints to the training workflows (e.g. after Step 3, verify loss is decreasing and KL is bounded before proceeding, with a fix-and-retry loop) to lift workflow clarity.

Move the inline "Core API Reference" and "Common Issues and Solutions" sections into the existing references/api-reference.md and references/troubleshooting.md files and link them from the body with "See [references/api-reference.md](...)".

Trim the marketing-style "Key Features" and "When to Use" bullets and the duplicated ASCII architecture diagram to reduce token overhead and redundancy.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient concrete code, configs, and commands, but at ~430 lines it carries some padding (the marketing-style "Key Features" bullets and an ASCII architecture diagram restated in references) and minor prose redundancy that could be tightened.	2 / 3
Actionability	It provides copy-paste-ready executable guidance throughout: install scripts, full YAML configs, a runnable reward function, a complete custom loss class, and concrete launch commands for local and SLURM execution.	3 / 3
Workflow Clarity	The three workflows are numbered with prerequisite checklists, but long-running/batch RL training lacks explicit validate→fix→retry feedback gates (e.g. a checkpoint verifying loss/entropy/KL health before continuing), which caps this dimension at 2.	2 / 3
Progressive Disclosure	The body keeps "Core API Reference" and "Common Issues and Solutions" inline and never links the existing references/ bundle files, whose api-reference.md even duplicates the architecture overview — content that should be split is inline and navigation to the bundles is missing.	2 / 3
	Total	9 / 12 Passed

Description

75%

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 solid, complete description with an explicit use-trigger and a clearly distinct niche, weakened only by a slightly vague leading verb and a few missing common trigger-term variations.

Suggestions

Replace the vague leading verb "Provides guidance for" with concrete actions, e.g. "Train and experiment with PyTorch-native agentic RL algorithms (GRPO, DAPO, SAPO) using torchforge".

Add common natural-language trigger terms users would actually say, such as "reinforcement learning", "GRPO", "policy gradient", or "RL fine-tuning", alongside the framework names.

Dimension	Reasoning	Score
Specificity	The description names the domain ("PyTorch-native agentic RL using torchforge") and capability areas ("clean RL abstractions, easy algorithm experimentation, scalable training"), but the leading verb "Provides guidance" is vague rather than a concrete action like "Train" or "Implement".	2 / 3
Completeness	It explicitly answers both "what" ("Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms") and "when" via an explicit "Use when you want..." trigger clause.	3 / 3
Trigger Term Quality	It surfaces relevant terms ("RL abstractions", "algorithm experimentation", "scalable training", "Monarch", "TorchTitan"), but leans on niche framework names and omits common variations a user would naturally say like "GRPO", "reinforcement learning", or "policy gradient".	2 / 3
Distinctiveness Conflict Risk	The combination of "torchforge", "PyTorch-native agentic RL", "Monarch", and "TorchTitan" carves out a very specific niche with distinct triggers that are unlikely to conflict with other skills.	3 / 3
	Total	10 / 12 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

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

Reviewed: about 12 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.