CtrlK
BlogDocsLog inGet started
Tessl Logo

modal-ml-training

Disconnect-safe patterns for long-running ML training on Modal serverless GPU. Covers the deploy+spawn pattern (survives laptop shutdown/SSH disconnect), checkpoint-resume for preemption recovery, PyTorch/CUDA version pinning, volume reload/commit discipline, and batch parameter sweeps. Use for any training job >30 min where losing progress is expensive. Complements the broader `modal-serverless-gpu` skill.

71

Quality

87%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

82%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 a dense, actionable pattern guide with executable code, a useful pitfalls table, and clear sequencing for the core disconnect-safe workflow. It is held back only by minor verbosity and the absence of explicit error-recovery feedback loops in the main flows.

Suggestions

Add an explicit validate→fix→retry feedback loop to the deploy+spawn workflow (e.g., verify the function appears in `modal app list` after deploy before spawning, and what to do if it does not).

Tighten framing prose such as the "NEVER use `modal run --detach`" preamble and the PyTorch-pinning intro to rely more on the WRONG/RIGHT code contrast already shown.

Consider splitting the GPU Selection Guide and Cost Estimation Template into a short reference file so the core training patterns stay front-and-center, improving progressive disclosure for this 200+ line skill.

DimensionReasoningScore

Conciseness

Mostly lean and code-focused with non-obvious domain explanations that earn their place (e.g., CUDA driver mismatch, volume race conditions), though a few framing sentences and the "NEVER use" preamble could be trimmed slightly.

4 / 5

Actionability

Provides copy-paste-ready code and commands for each taught mechanism — deploy, spawn, checkpoint save/load, volume reload/commit/read, parameterized sweeps, and monitoring — with the only placeholders being the user's own domain-specific training logic.

5 / 5

Workflow Clarity

The deploy+spawn flow is a clearly sequenced 3-step process and the batch section includes verification ("Use `modal app list` to verify task count"), but the main flows lack explicit validate→fix→retry feedback loops, leaving minor checkpoint gaps.

4 / 5

Progressive Disclosure

Well-organized into clearly headed sections (When to Use, patterns, GPU guide, pitfalls table) with no nested or buried references and a single signaled pointer to the broader related skill, though all content is inlined in one file with no deeper reference material.

4 / 5

Total

17

/

20

Passed

Description

92%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 specific, complete, and well-differentiated, with concrete pattern names and an explicit use-when clause. Voice is correctly third person throughout. The only minor gap is trigger-term synonym coverage.

DimensionReasoningScore

Specificity

Lists multiple concrete patterns — "deploy+spawn pattern", "checkpoint-resume for preemption recovery", "PyTorch/CUDA version pinning", "volume reload/commit discipline", "batch parameter sweeps" — giving comprehensive, specific coverage rather than vague language.

5 / 5

Completeness

Explicitly answers both what ("Disconnect-safe patterns for long-running ML training on Modal serverless GPU. Covers...") and when ("Use for any training job >30 min where losing progress is expensive") with concrete trigger phrases.

5 / 5

Trigger Term Quality

Natural phrases like "ML training", "training job >30 min", "laptop shutdown/SSH disconnect", and "losing progress is expensive" map to what a user would say, but synonyms and broader variations are limited compared to a fully comprehensive trigger set.

4 / 5

Distinctiveness Conflict Risk

Carves a clear niche (disconnect-safe long-training) and explicitly distinguishes itself: "Complements the broader `modal-serverless-gpu` skill", minimizing conflict risk with the related general skill.

5 / 5

Total

19

/

20

Passed

Validation

93%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation15 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

Total

15

/

16

Passed

Repository
synthetic-sciences/openscience
Reviewed

Table of Contents

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.