The first widely-replicated open reproduction of R1-style "zero RL" — RL with verifiable rewards directly on base models, no SFT cold start. The repo went live in January 2025, weeks after DeepSeek-R1, showing long chain-of-thought and self-verification emerging on a 7B base with only 8K MATH examples; SimpleRL-Zoo (the paper) then systematized zero RL across 10 base-model series (Qwen2.5 0.5B–32B, Llama-3.1-8B, DeepSeek/Mistral bases), releasing all checkpoints (3.9K★).

A heavily cited reference for the RL-for-reasoning wave and the anchor of Junxian He's hkust-nlp RL line (with Dr.Kernel, B-STaR, and Laser). Trained with veRL-based infrastructure; all models are RL-tuned open bases, not from-scratch pretrains.

Model Details

Base model qwen2.5

Variants

Name Parameters Notes
Qwen-2.5-7B-SimpleRL-Zoo zero-RL on Qwen2.5-7B base
Qwen-2.5-Math-7B-SimpleRL-Zero the original January 2025 release
Llama-3.1-8B-SimpleRL-Zoo zero-RL on Llama-3.1-8B base

Paper

reasoningpost-trainingopen-weightresearch