{"id":"verl-project-verl","name":"verl","af_score":36.5,"security_score":29.2,"reliability_score":35.0,"what_it_does":"verl is an open-source reinforcement learning (RL) training library for large language models (LLMs). It provides a flexible HybridFlow-style programming model to compose RL post-training dataflows (e.g., PPO/GRPO/ReMax/RLOO/REINFORCE++ and other recipes), and integrates with common LLM training/inference stacks (FSDP/FSDP2/Megatron-LM for training; vLLM/SGLang/HF Transformers for rollout generation).","best_when":"You need an extensible RL training framework for LLMs that can integrate multiple rollout backends and distributed training strategies (FSDP/Megatron/vLLM/SGLang), especially for large-scale RLHF-style post-training.","avoid_when":"You only need a simple model evaluation/inference API, or you cannot support PyTorch/distributed training and the associated engineering complexity.","last_evaluated":"2026-03-29T13:09:23.720933+00:00","has_mcp":false,"has_api":false,"auth_methods":[],"has_free_tier":false,"known_gotchas":["verl is a distributed RL training framework; agent-like automation must manage long-running jobs, cluster state, and checkpointing rather than stateless request/response flows.","Extensive integration with external backends (FSDP/Megatron/vLLM/SGLang) means failures may originate in those systems, and error semantics may be non-uniform."],"error_quality":0.0}