Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/axect/pytorch_template/pytorch-trainnpx skills add Axect/pytorch_template --skill pytorch-traingit clone --depth 1 https://github.com/Axect/pytorch_templateWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00127 | $0.02898 |
| Opus 5 | $0.00063 | $0.01449 |
| Sonnet 5 | $0.00025 | $0.00580 |
| Haiku 4.5 | $0.00013 | $0.00290 |
Grade A, and why
pytorch-train scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 300 lines — stays where its author put it; the contents beside it link to each section on GitHub.
pytorch-train
Train models through this template's full pipeline: config creation, HPO, best-param extraction, and final training.
Usage
/pytorch-train [phase] [options]
Phases (auto-detected from user intent if omitted):
full— Complete pipeline (Phases 1-5)config— Create config files only (Phases 1-2)hpo— Run HPO (Phase 3, assumes configs exist)extract— Extract best params from completed HPO (Phase 4)train— Run final training with best.yaml (Phase 5)
Phase 1: Gather Requirements
Before creating any files, confirm these with the user:
| Item | Example | Notes |
|---|---|---|
| Project name | SolarFlux, WavePredict |
Used in project: field and directory names |
| Version | v0.3, v1.32 |
Determines config subdirectory |
| Model name | fluxnet, wavenet, mlp |
File prefix and net: path |
| Task type | regression / classification | Determines criterion, metric direction |
| Model module path | model.MLP, recipes.regression.model.MLP |
Importlib path for net: field |
| net_config | {nodes: 64, layers: 4} |
Architecture hyperparameters |
| Data loader | existing load_data() or needs new one |
Recipe scaffolding if needed |
| Device | cuda:0 (default if available) |
GPU index |
Task Type Defaults
| Setting | Regression | Classification |
|---|---|---|
criterion |
torch.nn.MSELoss |
torch.nn.CrossEntropyLoss |
criterion_config |
{} |
{} |
metric (HPO) |
val_loss |
val_loss |
direction (HPO) |
minimize |
minimize |
Pre-flight Check (Before HPO)
After creating configs, always run preflight before HPO:
python -m cli preflight configs/<DIR>/<model>_run.yaml --device cuda:0
Preflight runs 1 batch forward+backward and checks:
- Import paths and device availability
- Semantic validation (upper_bound >= total_steps, lr > 0, etc.)
- Model/optimizer/scheduler/criterion instantiation
- Data loading via the config's
datafield - Forward pass (output shape vs target shape)
- Backward pass (gradient NaN/Inf detection, grad norm)
- GPU memory estimate
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 300 lines · 127 tokens per session scan A 2028f2104e07
pytorch-train is a skill published in the GitHub repository Axect/pytorch_template (10 stars, last pushed 3mo ago), licensed MIT. It adds 127 tokens to every session and 2,898 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
perforatedai
Expert in PerforatedAI library for adding artificial dendrites to PyTorch neural networks. Triggers: 'Perforate my model' (start interactive setup), 'debug my perforated model' (debug/optimize existing integration), 'load my perforated model for inference' (deploy trained models). Also use when: debugging dendrite…
addressing-pr-review-comments
Address all valid review comments on a PR for the current branch in the streamlit/streamlit repo. Covers both inline review comments and general PR (issue) comments. Use when a PR has reviewer feedback to address, including code changes, style fixes, and documentation updates.
generating-changelog
Generates polished website release notes between two git tags for docs.streamlit.io. Use when preparing a new Streamlit release or reviewing changes between versions.
understanding-streamlit-architecture
Explains Streamlit's internal architecture including backend runtime, frontend rendering, and WebSocket communication. Use when debugging cross-layer issues, understanding how features work end-to-end, planning architectural changes, or onboarding to the codebase. Covers ForwardMsg/BackMsg protocol, script rerun…
debugging-streamlit
Debug Streamlit frontend and backend changes using make debug with hot-reload. Use when testing code changes, investigating bugs, checking UI behavior, or needing screenshots of the running app.
fixing-flaky-e2e-tests
Diagnose and fix flaky Playwright e2e tests. Use when tests fail intermittently, show timeout errors, have snapshot mismatches, or exhibit browser-specific failures.