Skills
.Work. You
Rest

verl-agent-training

Community

One command trains a domain PPO agent end-to-end.

Software Engineering#agent training#qwen#huggingface#verl#dataset conversion#ppo training#gpu automation
AuthorCodekiing
Version1.0.0
Installs0

System Documentation

What problem does it solve?

It eliminates the manual, multi-step setup required to train a domain-specific PPO agent by automating environment preparation, dataset acquisition/conversion, and the VERL PPO training run with guardrails.

Core Features & Use Cases

  • End-to-end VERL PPO training pipeline: Clones VERL, installs a fixed compatible dependency set, prepares dataset parquet files, then launches python3 -m verl.trainer.main_ppo.
  • Interactive dataset selection by agent type: Uses your provided AGENT_TYPE to search and choose an appropriate Hugging Face dataset, then converts it into VERL-ready train.parquet and test.parquet.
  • Crash recovery via “immediate repair” rules: Provides deterministic fallback steps for common failures, including OOM mitigation by reducing response length and micro-batch sizes.

Quick Start

Run this Skill in a blank directory by first saving the provided script as run_verl_skill.sh, then executing chmod +x run_verl_skill.sh && bash run_verl_skill.sh after setting AGENT_TYPE to your desired domain (e.g., math).

Dependency Matrix

Required Modules

None required

Components

Standard package

💻 Claude Code Installation

Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.

Please help me install this Skill:
Name: verl-agent-training
Download link: https://github.com/Codekiing/Auto_VeRL_PPO_Skill/archive/main.zip#verl-agent-training

Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
View Source Repository

Agent Skills Search Helper

Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.