DeepExperienceOfficial·9 Agent Skills Included

HyperEyes

Parallel multimodal search training with efficiency-aware reinforcement learning

Trains multimodal search models that ground and retrieve multiple image entities in a single parallel step. Eliminates redundant sequential tool calls, slow inference loops, and manual RL pipeline setup across Megatron, SGLang, and Ray. Includes guides for model integration, recipe migration, distributed debugging, code review, and bilingual documentation.

71|0

↗

npx skills add DeepExperience/HyperEyes --all -g -y

Available:

Tells the agent how the Relax training framework is structured, which coding and distributed-computing rules to follow, and which expert sub-skill to load for each task such as model integration, debugging hangs, or code review.

All Skills in This Repository (9)

Pure Emerald Level Indicators

📦 In Repo

DeepExperience

model-integration

Guide new model architecture integration into the Relax training pipeline.

git-commit

Generate Conventional Commits messages with structured markdown bodies.

creating-skills

Create structured SKILL.md definitions with valid metadata and operational instructions.

redaccel-to-relax

Migrates RedAccel reinforcement learning algorithms to Relax-compatible modules and interfaces.

doc-writer

Create bilingual VitePress documentation pages verified against Relax source code.

verl-to-relax

Convert verl reinforcement learning recipes to Relax framework implementations.

code-review

Review git changes for Python, PyTorch, and ML code quality issues.

debug-hang

Diagnose stalled Ray distributed training tasks by tracing blocked execution paths.

relax-dev-debug

Debug and validate Relax reinforcement learning workflows on Ray clusters.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install HyperEyes?▼

Run `npx skills add DeepExperience/HyperEyes --all -g -y` in your terminal to install all skills in this suite globally.

What problem does HyperEyes solve?▼

It trains multimodal search models to ground and retrieve multiple image entities in parallel instead of slow sequential tool calls, cutting tool-call rounds by over 5x while improving accuracy.

How to migrate verl or RedAccel recipes to Relax?▼

Load the verl-to-relax or redaccel-to-relax skill, which maps reward functions, tool environments, datasets, and launch scripts to the Relax format step by step.

How to debug a hung distributed training job?▼

The debug-hang skill walks through checking Ray cluster status, collecting py-spy stack traces, and tracing actor dependency chains to find the root cause.

Do I need RL framework experience to use these skills?▼

Basic familiarity helps, but each skill gives step-by-step instructions, decision trees, and code templates so the agent handles the complex framework details for you.