SKU RLE-1402 · Sold by External

Search-R1

Product specifications

SKURLE-1402
Data typeRL environments
Volume retrieval-augmented QA training framework
Size on diskNot published by source
FormatPython (built on veRL; installed from source via conda/pip)
Access modelPUBLIC LICENSE
PricingFree · open-source license
Quality score
LicenseApache-2.0
Search-R1 is an RL framework (built on veRL) for training reasoning-and-searching LLMs that emit <search>query</search>, receive retrieved <information>, and produce a final <answer>. The search engine (local sparse/dense retrievers over a Wikipedia corpus, or online search) acts as the environment; it uses retrieved-token masking for stable RL and an outcome-based reward. Supports PPO, GRPO, and REINFORCE across Llama3 and Qwen2.5 models; trained and evaluated across 7 QA benchmarks. It is a training framework rather than a fixed-count environment catalog.