ClawHub

mhc-algorithm

v0.1.0

Implement mHC (Manifold-Constrained Hyper-Connections) for stabilizing deep network training. Use when implementing residual connection improvements with dou...

⭐ 0· 34·0 current·0 all-time
by@lnj22
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
✓
Purpose & Capability
The name/description (mHC for stabilizing deep nets) aligns with the contents: PyTorch code snippets, Sinkhorn projection, and GPT integration patterns. The only external dependency suggested (torch, einops, numpy) is appropriate for the stated goal.
✓
Instruction Scope
SKILL.md contains concrete implementation guidance, example code, and algorithm notes. Instructions are confined to model-code concerns (tensor shapes, Sinkhorn iterations, wrapping layers) and do not instruct reading arbitrary files, accessing environment variables, or contacting external endpoints beyond citing arXiv links.
ℹ
Install Mechanism
No install spec is embedded; the doc recommends 'pip install torch einops numpy'. This is expected for a PyTorch implementation but be aware 'pip install torch' can be large and platform-specific (CUDA variants). There are no downloads from untrusted URLs or archive/extract steps.
✓
Credentials
The skill requests no environment variables, credentials, or config paths. All required resources are typical Python packages needed to run the examples.
✓
Persistence & Privilege
The skill is instruction-only, does not request 'always: true', and does not instruct changing agent-wide configuration or storing credentials. It does not grant persistent or elevated privileges.
Assessment
This skill appears internally consistent and focused on implementing mHC in PyTorch. Before using: (1) Review the code snippets and references to ensure they fit your model and framework versions; (2) install PyTorch via the official channel appropriate for your OS/GPU (avoid arbitrary wheel URLs); (3) run the examples in an isolated environment (virtualenv/container) because mHC multiplies memory usage by num_streams; (4) confirm the referenced paper(s) if you need research provenance. If you need broader audits (license, benchmark results, or GPU/CUDA compatibility), request the author's complete implementation or test on a small toy model first.

Like a lobster shell, security has layers — review code before you run it.

latestvk975qnc3jz6sfc7wxwvwqyx95x84v8wq

License

MIT-0
Free to use, modify, and redistribute. No attribution required.
Termshttps://spdx.org/licenses/MIT-0.html

Comments