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

66%

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 ./backend/cli/skills/ml-training/torchforge/SKILL.md

Quality

Content

46%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 rich and actionable with strong code/config examples, but it undermines its own progressive-disclosure structure by inlining large blocks that duplicate the reference files rather than linking to them, and its workflows lack explicit validation checkpoints for batch training. Conciseness and progressive disclosure are the weakest dimensions, both driven by the same duplication problem.

Suggestions

Replace the inlined troubleshooting section and architecture/API blocks with pointers to references/troubleshooting.md and references/api-reference.md, keeping only the minimal quick-start content in SKILL.md.

Add explicit validation/feedback checkpoints to each workflow (e.g., 'verify W&B loss is decreasing and KL stays bounded before continuing; if entropy collapses, raise beta and restart') so batch training has a validate-then-proceed loop.

Tighten the opening paragraph to avoid restating the frontmatter description, and mark illustrative-only snippets like `inputs = [{"tokens": torch.Tensor}]` as pseudo-type hints so they are not mistaken for executable code.

Dimension	Reasoning	Score
Conciseness	The ~430-line body duplicates material already present in references/ (architecture diagram, YAML config, launch commands, and a full troubleshooting section mirror api-reference.md and troubleshooting.md), and restates the description in the opening paragraph, which is noticeably padded rather than pointed to.	2 / 5
Actionability	Provides concrete, mostly copy-paste-ready guidance (install scripts, full YAML configs, a runnable reward function, a complete loss class, and launch commands), with only minor illustrative-only snippets like `inputs = [{"tokens": torch.Tensor}]` keeping it off 5.	4 / 5
Workflow Clarity	Workflows are sequenced (GRPO Steps 1-4, Custom Loss Steps 1-2, Distributed launch) but lack explicit validation/verification gates, and for a long-running batch training operation the missing 'validate then proceed' feedback loop caps the score at 3 per the batch-operation rule.	3 / 5
Progressive Disclosure	Bundle reference files exist (references/api-reference.md, references/troubleshooting.md) but the body never links to them, instead inlining the very content (full troubleshooting issues, architecture diagram, API reference) that belongs in those separate files.	2 / 5
	Total	11 / 20 Passed

Description

87%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 well-constructed: it states a concrete purpose, names the specific library and stack, and provides an explicit 'Use when...' trigger clause covering multiple scenarios. Its main limitation is slightly generic phrasing ('Provides guidance') and missing synonym/extension coverage, which keep trigger_term_quality and specificity at 4 rather than 5.

Dimension	Reasoning	Score
Specificity	Names the domain and library ('PyTorch-native agentic RL using torchforge') and lists several specific capability drivers ('clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan'), with only the generic 'Provides guidance' verb keeping it off 5.	4 / 5
Completeness	Explicitly answers both 'what' ('Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms') and 'when' ('Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training...') with concrete trigger phrases.	5 / 5
Trigger Term Quality	Includes natural terms users would say ('PyTorch', 'agentic RL', 'RL abstractions', 'algorithm experimentation', 'Monarch', 'TorchTitan'), but lacks synonyms and file-extension variants that would push it to 5.	4 / 5
Distinctiveness Conflict Risk	Occupies a clear niche (PyTorch-native agentic RL via torchforge) with distinctive triggers (Monarch, TorchTitan, GRPO/DAPO/SAPO) that minimize overlap with generic training skills.	5 / 5
	Total	18 / 20 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/torchforge/SKILL.md
Commit: 04797e6

Reviewed: 3 days 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.