#5353rd of 53 in Training and fine-tuning
search-r1
RL framework that trains LLMs to interleave reasoning with search engine calls
- Stars
- 5.6k
- License
- Apache-2.0
- Last commit
- Nov 2025
- Language
- Python
4,489 on 21 Apr 20265,574 on 10 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Search-R1 is a training pipeline built on veRL that uses rule-based outcome rewards to teach base LLMs, such as Llama-3.2-3B and Qwen2.5-7B, to reason and call a search engine over multiple turns. It supports PPO, GRPO and reinforce, with retrieval from local BM25, dense retrievers with flat or ANN indexes, or online search APIs. The search engine runs as a separate HTTP server that training calls.
Who it is for: Researchers training search-calling reasoning LLMs with RL
Strengths
- Supports PPO, GRPO and reinforce training methods
- Pluggable retrievers: BM25, dense flat/ANN, online search APIs, rerankers
- Multinode training documented for 30B+ models
- Two papers and public wandb logs for reproducing experiments
Weaknesses
- Pinned older stack: Python 3.9, torch 2.4.0, vllm 0.6.3
- Needs two separate conda environments plus a running retrieval server
- No Docker or compose setup provided
- Quick start downloads a Wikipedia corpus and index before training
What it needs
- GPU required
- Needs veRL, vLLM 0.6.3, PyTorch 2.4.0, flash-attn, faiss-gpu 1.8.0, pyserini, FastAPI, wandb
- Models: Llama-3.2-3B, Qwen2.5-3B, Qwen2.5-7B
- port 8000
Quick answers
Is search-r1 open source?
Yes. Its code is on GitHub under the Apache-2.0 license.
Does search-r1 need a GPU?
Yes, it needs a GPU.
Also in Training and fine-tuning
See all 53| Rank | Project | Score |
|---|---|---|
| 1 | unslothDesktop and web app to run and fine-tune local models | 88 out of 100 |
| 2 | peftPython library for LoRA and other parameter-efficient fine-tuning of Transformers models | 88 out of 100 |
| 3 | deepspeedPyTorch library for distributed training of very large models with ZeRO | 86 out of 100 |
| #4 | verlReinforcement learning training library for LLM post-training with PPO and GRPO | 86 out of 100 |
| #5 | trlPost-training library with SFT, DPO, GRPO and KTO trainers | 86 out of 100 |