agentmods.dev

jepa-pytorch-development

jepa-pytorch-development is a skill for Claude Code, Codex from CHENyiru3/AI-Skills-Collections. It costs 77 tokens per session (1,664 once invoked), scanned A, original, MIT.

A PyTorch development guide for JEPA models, which learn useful hidden representations by predicting missing or future information. It covers image, video, view-invariant, and action-conditioned variants.

In plain words
What is it for?
It is for designing, implementing, adapting, or debugging JEPA representation learners and latent world models in PyTorch.
Why use it?
It helps prevent mixing incompatible model parts, such as masking rules, teacher networks, prediction targets, and methods for avoiding collapsed representations.

Skill for Claude CodeCodex

Written for Claude Code and Codex: shipped in a Claude Code plugin, but also agents/openai.yaml present.

Part of the ai-ml plugin — 14 skills shipped together

not rated 2repo changed 3d ago A scan 77 tokens original MIT

Good fit It is for designing, implementing, adapting, or debugging JEPA representation learners and latent world models in PyTorch.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/chenyiru3/ai-skills-collections/jepa-pytorch-development
View source ↗ CHENyiru3/AI-Skills-Collections
Install

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.

Any agent
npx skills add CHENyiru3/AI-Skills-Collections --skill jepa-pytorch-development
Clone the repo
git clone --depth 1 https://github.com/CHENyiru3/AI-Skills-Collections

Made for: Claude Code, Codex.

Or install ai-ml, the plugin that ships this one along with the rest of its 14 skills.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for jepa-pytorch-development

README.md
[![agentmods](https://agentmods.dev/badge/skills/chenyiru3/ai-skills-collections/jepa-pytorch-development/github.svg)](https://agentmods.dev/skills/chenyiru3/ai-skills-collections/jepa-pytorch-development)
Your own site
<a href="https://agentmods.dev/skills/chenyiru3/ai-skills-collections/jepa-pytorch-development"><img src="https://agentmods.dev/badge/skills/chenyiru3/ai-skills-collections/jepa-pytorch-development/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for jepa-pytorch-development

Your own site · 80×15
<a href="https://agentmods.dev/skills/chenyiru3/ai-skills-collections/jepa-pytorch-development"><img src="https://agentmods.dev/badge/skills/chenyiru3/ai-skills-collections/jepa-pytorch-development.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 77 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,664 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What 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.

ModelPer sessionOnce invoked
Fable 5.1 $0.00077 $0.01664
Opus 5.5 $0.00031 $0.00666
Sonnet 5.5 $0.00015 $0.00333
Haiku 4.5 $0.00008 $0.00166

Measured 3d ago against content hash 3ea32850d2b7, method: parsed. Prices are Anthropic first-party input rates as of 2026-10-07, from the pricing page.

Security

Grade A, and why

jepa-pytorch-development 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 3d ago.

The scan reads SKILL.md. This mod also ships 1 executable file (scripts/smoke_masked_jepa.py), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

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.

skills-market/ai-ml/deep-learning/jepa-pytorch-development/SKILL.md · 83 lines

How it starts

The opening of the file, as written. The whole thing — 83 lines — stays where its author put it; the contents beside it link to each section on GitHub.

JEPA development in PyTorch

Build around an explicit prediction task and information boundary. Choose a coherent recipe before borrowing components. A JEPA name alone does not specify masking, an EMA teacher, a predictor, or the collapse-prevention mechanism.

Select the recipe

Inspect the user's code, configuration, pretrained checkpoint, compute budget, and downstream task. Infer routine choices; ask only when an unresolved choice changes the model's meaning. Distinguish reproduction, a deliberate variant, debugging, and transfer to a new modality.

Task Starting family Read when needed
Predict missing image/video features I-JEPA / V-JEPA / V-JEPA 2: visible-token encoder, positional predictor, stopped EMA targets Masked EMA
Learn dense video features at multiple depths V-JEPA 2.1: visible/masked prediction with aligned intermediate targets Dense/deep supervision
Learn view-invariant representations without a teacher LeJEPA: shared encoder/projector, invariance plus SIGReg Teacher-free learning
Predict future states conditioned on actions Frozen-encoder V-JEPA 2-AC, or explicitly regularized joint training World models

For attribution or faithful reproduction, consult the commit-pinned repository map. It distinguishes observed source behavior from engineering recommendations. Do not claim benchmark reproduction from a toy smoke test.

Define the contracts before implementing the loss

Write these in the project's existing config/design conventions:

  • Input units/layout, tokenizer geometry, flattening order, padding and validity. For a standard video grid, N = T_tokens * H_tokens * W_tokens; assert actual tokenizer output agrees.
  • Context/target indices in original coordinates, with batch, view, mask, time, and layer axes named. Specify paired versus Cartesian multi-mask expansion and prediction-target alignment.
  • Allowed online/predictor inputs, permitted overlap, target normalization, prediction dimensions, reduction denominators, and component weights.
  • Gradient ownership: optimized, frozen, or EMA-updated modules; teacher train/eval behavior and mutable-buffer policy.
  • Augmentations and split units. Preserve correspondence within a masked-prediction pair; use the chosen view policy for invariance. Split trajectories by episode/entity where appropriate.

Read the full file on GitHub · 83 lines

Files

What ships with it

6 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.

Changes

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.

  1. 3d ago Changed · +2 lines 3ea32850d2b7
  2. 10d ago First seen · 81 lines · 77 tokens per session scan A b65ca92119b7

Subscribe to this mod's changes

jepa-pytorch-development is a skill published in the GitHub repository CHENyiru3/AI-Skills-Collections (2 stars, last pushed 3d ago), licensed MIT. It adds 77 tokens to every session and 1,664 once invoked, about $0.0003 per session on Opus 5.5. 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 2026-11-04.

Data and AIScientific
Related

jepa-pytorch-development compared with skills from other repositories

jepa-pytorch-development next to 6 skills from other repositories
ModTypeStarsLast pushLicenceScanTokensWorks with
jepa-pytorch-developmentCHENyiru3/AI-Skills-CollectionsSkill23d agoMITA77Claude Code, Codex
arboretoK-Dense-AI/scientific-agent-skillsSkill57k5d agoMITA55Claude Code, Codex
deepspot-mK-Dense-AI/scientific-agent-skillsSkill57k5d agoMITA70Claude Code
torchdrugK-Dense-AI/scientific-agent-skillsSkill57k5d agoMITA64Claude Code
pyhealthK-Dense-AI/scientific-agent-skillsSkill57k5d agoMITA84Claude Code, Codex
nemo-mbridge-perf-expert-parallel-overlapNVIDIA/skillsSkill4.0kyesterdayApache-2.0A56Claude Code
evo2aipoch/open-scienceSkill5.5ktodayApache-2.0A83Claude Code, Codex