跳转到主要内容
此内容尚未提供您的语言版本,正在以英文显示。

Verl Rl Training

技能 活跃

Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.

目的

To enable users to implement advanced LLM post-training techniques like RLHF, GRPO, and PPO at scale using the verl library.

功能

  • Guidance for RLHF, GRPO, PPO, and other RL algorithms
  • Support for large-scale LLM post-training
  • Flexible infrastructure backend configurations
  • Detailed installation and quick start examples
  • Comprehensive configuration reference

使用场景

  • Implementing RLHF for LLM fine-tuning
  • Training LLMs with GRPO for reasoning tasks
  • Scaling PPO training for large language models
  • Leveraging flexible backends like FSDP, Megatron-LM, and vLLM

非目标

  • Megatron-native training (recommends other tools)
  • PyTorch-native abstractions with Monarch (recommends other tools)
  • Simple SFT/DPO (recommends other tools)

工作流

  1. Prepare Dataset
  2. Define Reward Function
  3. Create Training Config
  4. Launch Training
  5. Monitor and Validate

实践

  • Reinforcement Learning
  • LLM Post-Training
  • Distributed Systems

先决条件

  • GPU cluster with 8+ GPUs (H100 recommended for math tasks)
  • Dataset in parquet format with 'prompt' and 'reward_model' columns
  • Base model from HuggingFace Hub
  • Install Megatron-LM bridge (for Megatron workflow)

Trust

  • warning:Issues AttentionIn the last 90 days, 17 issues were opened and 4 were closed, indicating a low closure rate and potentially slow maintainer response.

安装

npx skills add davila7/claude-code-templates

通过 npx 运行 Vercel skills CLI(skills.sh)— 需要本地安装 Node.js,以及至少一个兼容 skills 的智能体(Claude Code、Cursor、Codex 等)。前提是仓库遵循 agentskills.io 格式。

质量评分

95 /100
about 18 hours ago 分析

信任信号

最近提交about 20 hours ago
星标27.2k
许可证MIT
状态
查看源代码

类似扩展

Verl Rl Training

99

Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.

技能
Orchestra-Research

Fine Tuning With Trl

96

Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.

技能
Orchestra-Research

Openrlhf Training

99

High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.

技能
Orchestra-Research

Slime Rl Training

98

Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. Use when training GLM models, implementing custom data generation workflows, or needing tight Megatron-LM integration for RL scaling.

技能
Orchestra-Research

Openrlhf Training

97

High-performance RLHF framework with Ray+vLLM acceleration. Use for PPO, GRPO, RLOO, DPO training of large models (7B-70B+). Built on Ray, vLLM, ZeRO-3. 2× faster than DeepSpeedChat with distributed architecture and GPU resource sharing.

技能
davila7

Torchforge

96

Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.

技能
Orchestra-Research