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

65%

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

65%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The content is strongly actionable with concrete configs and executable code, but it underuses progressive disclosure by duplicating reference material inline instead of linking the existing bundle files, and its training workflows lack embedded validation feedback loops.

Suggestions

Link the existing bundle files from the body: replace the inline 'Common Issues and Solutions' section with 'See [troubleshooting.md](references/troubleshooting.md)' and defer architecture detail to '[api-reference.md](references/api-reference.md)' to avoid duplication.

Add explicit validation checkpoints with feedback loops to the training workflows, e.g. after launching: 'Verify loss is finite and KL is bounded; if NaN/OOM, reduce batch_size or seq_len and relaunch.'

Remove the duplicated inline troubleshooting section and the repeated architecture diagram (both already in the bundle files) to recover roughly 80 lines of token budget.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient and code-dense, but it duplicates content that already exists as bundle files — the 'Common Issues and Solutions' section mirrors references/troubleshooting.md and the architecture diagram is repeated from references/api-reference.md — adding roughly 80 lines that could be trimmed.	2 / 3
Actionability	Provides copy-paste-ready guidance throughout: install/verify commands, complete YAML configs, a full executable CustomLoss class, a concrete reward function, and SLURM/multi-GPU launch commands.	3 / 3
Workflow Clarity	Workflows are clearly sequenced (Step 1-4) with prerequisite and monitoring checklists, but the batch training workflows lack explicit validate->fix->retry feedback loops, and recovery guidance is split into a separate troubleshooting section rather than embedded as checkpoints.	2 / 3
Progressive Disclosure	The body is well-sectioned but never links the provided bundle files (references/api-reference.md, references/troubleshooting.md); instead it duplicates their architecture and troubleshooting content inline, so content that should be separate lives in the main file.	2 / 3
	Total	9 / 12 Passed

Description

65%

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 has an explicit 'Use when...' trigger and a distinctive niche, but it is weakened by the vague verb 'Provides guidance for', second-person phrasing, and trigger terms that lean on technical jargon while omitting common RL variations.

Suggestions

Replace 'Provides guidance for' with concrete actions the skill performs, e.g. 'Configures and implements GRPO/DAPO/SAPO loss functions and launches distributed PyTorch-native RL training with torchforge.'

Rewrite the trigger in third person to remove 'you want', e.g. 'Use when training reasoning models with GRPO, implementing custom RL losses, or scaling RL training with Monarch and TorchTitan.'

Add natural trigger terms users would actually say — 'reinforcement learning', 'RLHF', 'GRPO training', 'custom RL loss' — alongside the technical tool names.

Dimension	Reasoning	Score
Specificity	Names the domain ('PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms') but leads with the vague verb 'Provides guidance for' rather than concrete actions; the second-person 'Use when you want...' triggers the rubric's -1 voice penalty, dropping it from 2 to 1.	1 / 3
Completeness	Explicitly states what the skill does ('Provides guidance for PyTorch-native agentic RL... separating infra from algorithms') and when to use it via an explicit 'Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training...' clause, satisfying both what and when.	3 / 3
Trigger Term Quality	Includes relevant terms ('RL', 'PyTorch', 'scalable training', 'algorithm experimentation') but leans on jargon ('Monarch', 'TorchTitan', 'PyTorch-native agentic RL') that users may not naturally say, and omits common variations like 'GRPO', 'RLHF', or 'reinforcement learning'.	2 / 3
Distinctiveness Conflict Risk	Targets a clear niche anchored by specific tooling ('torchforge', 'Monarch', 'TorchTitan') that is unlikely to trigger for unrelated skills.	3 / 3
	Total	9 / 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/torchforge/SKILL.md
Commit: 52845c3

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.