dapo

Reproducing and studying RL algorithms for LLM agents, including GRPO, GSPO, DAPO, OPD, Search-R1, ReTool, ALFWorld and beyond.

LightRFT: Light, Efficient, Omni-modal & Reward-model Driven Reinforcement Fine-Tuning Framework

Codebase of GRPO: Implementations and Resources of GRPO and Its Variants

InfraMind: Fine-tuning toolkit for training SLMs on Infrastructure-as-Code using GRPO/DAPO. Achieves 97.3% accuracy on IaC generation.

Open Ended Medical Reinforcement Learning

Unified reproductions, benchmarking & evaluation for perception-aware RLVR in VLMs. 11 methods · 25+ benchmarks · Official code of CGPO (ACM MM 2026 Oral).

An end-to-end framework for Agentic Reinforcement Learning.

Group Contrastive Policy Optimazation. Read the paper on arXiv: 👉 https://arxiv.org/abs/2510.07790

Exact finite-group identity behind GRPO reward standardization, unifying GRPO / Dr. GRPO / DAPO for RLVR and LLM reasoning. Paper + code.

Process-supervised RL for a multi-step reasoning agent - DAPO + a learned Process Reward Model training a Qwen3-8B Planner

Sync vs fully-async agentic RL on verl: multi-turn GRPO, long-tail rollout profiling, staleness ablations — quantifying when async pays off.

Unofficial PyTorch reproduction for DAPO: An Open-Source LLM Reinforcement Learning System at Scale.

Simple, zero-dependency tabular data manipulation and analysis for Python.

让 RL 工程里的坑不再只活在某个人的经验里。**当前是一个 demo ——** 只给结构、不给内容：36 个纯文本 Markdown 节点 + 一个只查自洽、不判对错的校验器 + 一个离线确定性检索器（无模型）。 | Make RL engineering pitfalls survive beyond one person's memory. **A demo —** structure, not content: 36 plain-text Markdown nodes, a consistency-only validator, and an offline deterministic matcher (no model).

Behavioral study of how SFT, DAPO and GRPO reshape mathematical reasoning distributions on Gemma-2-2B-IT.