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

58%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/ml-training/torchforge/SKILL.md

Quality

Content

60%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 with concrete, executable configs, code, and commands, and workflows are clearly sequenced with checklists in the flagship workflow. However, progressive disclosure fails badly: two substantial bundled reference files are never referenced while their content is duplicated inline, and install scripts are referenced that do not exist in the bundle.

Suggestions

Replace the inline 'Core API Reference' and 'Common Issues and Solutions' sections with pointers: 'API reference: See [references/api-reference.md](references/api-reference.md)' and 'Troubleshooting: See [references/troubleshooting.md](references/troubleshooting.md)'.

Either add the scripts/install.sh and scripts/install_rocm.sh files to the bundle or replace those references with inline pip/conda commands that can actually be run.

Add a validation step to Workflows 2 and 3, e.g. smoke-test a custom loss on a batch of dummy tensors before integrating it, and verify GPU allocation with a quick single-step dry run before launching SLURM jobs.

Dimension	Reasoning	Score
Conciseness	Prose is lean and most tokens are executable config/code, but the body inlines a 'Core API Reference' and a 'Common Issues and Solutions' section that duplicate the bundled references/api-reference.md and references/troubleshooting.md, adding ~150 lines of avoidable content. Not a 4 because this duplicated material is unnecessary given the bundle.	3 / 5
Actionability	Nearly all guidance is copy-paste executable: full YAML configs, complete Python loss and reward classes, and concrete launch commands. Not a 5 because './scripts/install.sh' and './scripts/install_rocm.sh' are referenced but no scripts/ directory exists in the bundle, and the custom-loss integration snippet uses undefined variables.	4 / 5
Workflow Clarity	All three workflows present clear numbered sequences, and Workflow 1 includes a prerequisites checklist plus explicit monitoring checkpoints ('Verify entropy is decreasing', 'Monitor KL divergence'). Not a 5 because Workflows 2 and 3 have no validation or verification checkpoints (e.g., smoke-testing a custom loss on a single GPU before launching distributed jobs).	4 / 5
Progressive Disclosure	The bundle provides references/api-reference.md (327 lines) and references/troubleshooting.md (409 lines), but the body never links to or mentions either file; instead it inlines duplicate 'Core API Reference' and 'Common Issues' sections. This matches the anchor for content that clearly belongs in separate files being inlined with references buried. Not a 3 because the reference files are not merely unclearly signaled — they are completely orphaned from the SKILL.md.	2 / 5
	Total	13 / 20 Passed

Description

57%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 answers both what and when explicitly and carves out a distinct niche by naming torchforge and its stack, but it is held back by second-person phrasing and generic action language ('Provides guidance'). Trigger terms lean technical, omitting natural phrases like 'reinforcement learning' or 'GRPO'.

Suggestions

Rewrite in third person: replace 'Use when you want clean RL abstractions...' with 'Use when working on agentic reinforcement learning, GRPO/DAPO training, or scalable RL with Monarch and TorchTitan.'

Replace the generic action 'Provides guidance for' with concrete capabilities, e.g. 'Implements RL algorithms (GRPO, DAPO, SAPO) decoupled from distributed infrastructure, and launches multi-GPU training with Monarch and TorchTitan.'

Add natural trigger synonyms users would say — 'reinforcement learning', 'RLHF', 'post-training', 'fine-tuning' — alongside the existing technical terms.

Dimension	Reasoning	Score
Specificity	The description names the domain ('PyTorch-native agentic RL') and enumerates covered areas ('clean RL abstractions, easy algorithm experimentation, or scalable training'), which sits at anchor 3, but the guideline-mandated third-person rule is violated by the second-person phrasing 'Use when you want', reducing the score by 1. It is not a 1 because the domain and scope are concrete rather than vague.	2 / 5
Completeness	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...') are explicitly present. Not a 5 because the when-clause frames user desires rather than concrete trigger phrases or situations.	4 / 5
Trigger Term Quality	Relevant keywords are present ('PyTorch-native agentic RL', 'training', 'Monarch', 'TorchTitan') but common natural variations users would actually say are missing ('reinforcement learning', 'GRPO', 'fine-tuning', 'RLHF', 'post-training'). Not a 4 because the included terms skew toward technical jargon over natural phrasing.	3 / 5
Distinctiveness Conflict Risk	Naming a specific library ('torchforge') and stack ('Monarch', 'TorchTitan') creates a clear niche with minimal conflict risk. Not a 5 because 'Provides guidance for... RL' still overlaps generically with other RL training libraries (verl, slime, miles).	4 / 5
	Total	13 / 20 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
metadata_version	'metadata.version' is missing	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	14 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/torchforge/SKILL.md
Commit: 4082a2e

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.