
understand-r1-zero
Understanding R1-Zero-Like Training: A Critical Perspective
Stars: 777

The 'understand-r1-zero' repository focuses on understanding R1-Zero-like training from a critical perspective. It provides insights into base models and reinforcement learning components, highlighting findings and proposing solutions for biased optimization. The repository offers a minimalist recipe for R1-Zero training, detailing the RL-tuning process and achieving state-of-the-art performance with minimal compute resources. It includes codebase, models, and paper related to R1-Zero training implemented with the Oat framework, emphasizing research-friendly and efficient LLM RL techniques.
README:
Zichen Liu*β , Changyu Chen*, Wenjun Li*, Penghui Qi*
Tianyu Pang, Chao Du, Wee Sun Lee, Min Lin
*Core Contributors, β Project Lead
π Updates β’ π Links β’ π TL;DR
- 21/03/2025: π We release our paper, models and codebase. Our R1-Zero training is implemented with πΎ Oat, a highly modular, research-friendly and efficient LLM RL framework.
-
Understanding R1-Zero-Like Training: A Critical Perspective
- π Paper
- π€ Models
- π Release tweet
-
There May Not Be Aha Moment in R1-Zero-like Training β A Pilot Study
- π Blog
- π» Code
- π Release tweet
-
OAT: A research-friendly framework for LLM online alignment
- π» Codebase
To understand R1-Zero-like training, we critically examine two core components: base models and reinforcement learning. We highlight our findings below.
- DeepSeek-V3-Base already exhibit "Aha moment".
- As the popular choice for R1-Zero-like training, Qwen2.5 base models demonstrate strong reasoning capabilities even without prompt templates: the average benchmark scores improve by ~60% (compared to the traditional 4-shot prompting)!
- GRPO leads to biased optimization! We propose a simple fix that improves token efficiency while maintaining reasoning performance, termed as Dr. GRPO (GRPO Done Right).
- In R1-Zero-like training, the template and the question set perform a duet to affect the RL dynamics
- (Left Plot) For Qwen2.5-Math-1.5B, a mismatched template (e.g., R1 template) in fact destroys the reasoning capabilities before RL reconstructing it. This makes the improvement impressive on the surface.
- (Middle Plot) However, if a template does not deviate from the pretraining distribution too far, even a small and completely o.o.d. question set (e.g., GSM8K) could induce the reasoning ability equally well, by reinforcing correct reasoning behaviors instead of infusing new knowledge.
- Beyond Qwen, Llama can also be RL-tuned from base models. In this case, domain-specific pretraining will improves RL ceiling.
- (Right Plot) GRPO can even make Llama with math knowledge "Aha" by increasing the output length; however, it is likely due to its length bias, which can be removed by Dr. GRPO.
Our analysis suggests a minimalist recipe for R1-Zero-like training:
We RL-tune Qwen2.5- Math-7B using the (unbiased) Dr. GRPO algorithm on MATH level 3-5 questions with the Qwen-Math template, and achieve state-of-the-art performance with only 27 hours compute on 8Γ A100 GPUs.
If you are interested in more details, please check out our paper!
We recommend a clean python==3.10
environment for development.
# Install vllm & oat, the LLM RL framework we developed r1-zero training on.
pip install vllm==0.7.2 && pip install oat-llm==0.0.9
# Install this package locally to use the math grader.
git clone [email protected]:sail-sg/understand-r1-zero.git && cd understand-r1-zero
pip install -e .
We implement R1-Zero training by extending Oat's Learner and Actor components. Please see train_zero_math.py for a step-by-step guide.
# Patch LD_LIBRARY_PATH to avoid dependency errors:
export LD_LIBRARY_PATH=$(python -c "import sysconfig; print(sysconfig.get_config_var('LIBDIR'))"):$LD_LIBRARY_PATH
# Run the experiment (tested on 8 x A100-40G) with Dr. GRPO:
# (change to `--critic_type grpo` for running GRPO)
python train_zero_math.py \
--critic_type drgrpo \
--gpus 8 \
--enable_prefix_caching \
--collocate \
--vllm_sleep \
--vllm_gpu_ratio 0.35 \
--gradient-checkpointing \
--flash-attn \
--bf16 \
--rnd-seed \
--learning_rate 0.000001 \
--lr_scheduler constant \
--num_ppo_epochs 1 \
--beta 0 \
--oracle_type reward \
--oracle math \
--pretrain Qwen/Qwen2.5-Math-1.5B \
--prompt_template r1 \
--zero-stage 2 \
--ref_offload \
--prompt_data ./datasets/train/math_12k \
--train_split train \
--input_key problem \
--output_key answer \
--max-train 9999999 \
--num_prompt_epoch 20 \
--prompt_max_length 1024 \
--num_samples 8 \
--temperature 1 \
--top_p 1 \
--generate_max_length 3000 \
--save_steps -1 \
--train_batch_size 128 \
--train_batch_size_per_device 1 \
--mini_train_batch_size_per_device 1 \
--rollout_batch_size 128 \
--rollout_batch_size_per_device 16 \
--pi_buffer_maxlen_per_device 128 \
--eval_batch_size 200 \
--eval_steps 16 \
--eval_temperature 0 \
--eval_generate_max_length 3000 \
--eval_data ./datasets/evaluation_suite \
--eval_input_key input \
--use-wb \
--wb-run-name qwen2.5-Math-1.5b-r1-zero \
--wb_project oat-zero
Please see here for more example scripts.
# Evaluate our models:
python evaluate_model.py --model_name sail/Qwen2.5-Math-7B-Oat-Zero
python evaluate_model.py --model_name sail/Qwen2.5-Math-1.5B-Oat-Zero
python evaluate_model.py --model_name sail/Llama-3.2-3B-Oat-Zero --template r1
# Evaluate baseline models:
python evaluate_model.py --model_name Qwen/Qwen2.5-Math-1.5B
python evaluate_model.py --model_name Qwen/Qwen2.5-Math-7B
python evaluate_model.py --model_name hkust-nlp/Qwen-2.5-Math-7B-SimpleRL-Zero
python evaluate_model.py --model_name PRIME-RL/Eurus-2-7B-PRIME-Zero
python evaluate_model.py --model_name Open-Reasoner-Zero/Open-Reasoner-Zero-7B
We provide a script to serve DeepSeek-V3-Base and DeepSeek-R1-Zero on k8s cluster.
# prerequisites:
# 1. download the model weights
# 2. starting a k8s job with sglang docker image "lmsysorg/sglang:v0.4.3.post2-cu125"
# start the server:
bash deploy_dpsk/serving.sh <model_name> <num_nodes>
Example of API call:
from openai import OpenAI
# MASTER_ADDR is the environment variable set by the k8s job
api_base = "http://{MASTER_ADDR}:30000/v1"
api_key = "EMPTY"
client = OpenAI(
api_key=api_key,
base_url=api_base,
)
# send requests to the server ...
Notes:
- Your k8s container should have environment variable
MASTER_ADDR
andMASTER_PORT
set. - Hardware requirements:
2 x 8 x H100/800/20
for FP8 and4 x 8 x A100/A800
for BF16. - Please refer to sglang's official tutorial for more details.
If you find our works useful for your research, please consider citing:
-
This paper:
@article{liu2025understanding, title={Understanding R1-Zero-Like Training: A Critical Perspective}, author={Zichen Liu and Changyu Chen and Wenjun Li and Penghui Qi and Tianyu Pang and Chao Du and Wee Sun Lee and Min Lin}, journal={arXiv preprint arXiv:2503.20783}, year={2025} }
-
Our blog that conducted the first investigation on the "Aha moment":
@misc{liu2025there, title={There May Not be Aha Moment in R1-Zero-like Training β A Pilot Study}, author={Zichen Liu and Changyu Chen and Wenjun Li and Tianyu Pang and Chao Du and Min Lin}, year={2025}, howpublished={\url{https://oatllm.notion.site/oat-zero}}, note={Notion Blog}, }
-
The training framework:
@misc{liu2025oat, title={OAT: A research-friendly framework for LLM online alignment}, author={Zichen Liu and Changyu Chen and Chao Du and Wee Sun Lee and Min Lin}, year={2025} howpublished={\url{https://github.com/sail-sg/oat}}, }
- This work is supported by Sea AI Lab for computing resources.
- The training codes are built on Oat, which employs vLLM, DeepSpeed and launchpad. We serve DeepSeek models using SGLang.
- The base models are from Qwen2.5-Math, Llama, and DeepSeek.
- We thank Qingfeng Lan for his time in thoroughly reviewing our code.
For Tasks:
Click tags to check more tools for each tasksFor Jobs:
Alternative AI tools for understand-r1-zero
Similar Open Source Tools

understand-r1-zero
The 'understand-r1-zero' repository focuses on understanding R1-Zero-like training from a critical perspective. It provides insights into base models and reinforcement learning components, highlighting findings and proposing solutions for biased optimization. The repository offers a minimalist recipe for R1-Zero training, detailing the RL-tuning process and achieving state-of-the-art performance with minimal compute resources. It includes codebase, models, and paper related to R1-Zero training implemented with the Oat framework, emphasizing research-friendly and efficient LLM RL techniques.

ichigo
Ichigo is a local real-time voice AI tool that uses an early fusion technique to extend a text-based LLM to have native 'listening' ability. It is an open research experiment with improved multiturn capabilities and the ability to refuse processing inaudible queries. The tool is designed for open data, open weight, on-device Siri-like functionality, inspired by Meta's Chameleon paper. Ichigo offers a web UI demo and Gradio web UI for users to interact with the tool. It has achieved enhanced MMLU scores, stronger context handling, advanced noise management, and improved multi-turn capabilities for a robust user experience.

stable-diffusion.cpp
The stable-diffusion.cpp repository provides an implementation for inferring stable diffusion in pure C/C++. It offers features such as support for different versions of stable diffusion, lightweight and dependency-free implementation, various quantization support, memory-efficient CPU inference, GPU acceleration, and more. Users can download the built executable program or build it manually. The repository also includes instructions for downloading weights, building from scratch, using different acceleration methods, running the tool, converting weights, and utilizing various features like Flash Attention, ESRGAN upscaling, PhotoMaker support, and more. Additionally, it mentions future TODOs and provides information on memory requirements, bindings, UIs, contributors, and references.

VideoTuna
VideoTuna is a codebase for text-to-video applications that integrates multiple AI video generation models for text-to-video, image-to-video, and text-to-image generation. It provides comprehensive pipelines in video generation, including pre-training, continuous training, post-training, and fine-tuning. The models in VideoTuna include U-Net and DiT architectures for visual generation tasks, with upcoming releases of a new 3D video VAE and a controllable facial video generation model.

RLAIF-V
RLAIF-V is a novel framework that aligns MLLMs in a fully open-source paradigm for super GPT-4V trustworthiness. It maximally exploits open-source feedback from high-quality feedback data and online feedback learning algorithm. Notable features include achieving super GPT-4V trustworthiness in both generative and discriminative tasks, using high-quality generalizable feedback data to reduce hallucination of different MLLMs, and exhibiting better learning efficiency and higher performance through iterative alignment.

ccprompts
ccprompts is a collection of ~70 Claude Code commands for software development workflows with agent generation capabilities. It includes safety validation and can be used directly with Claude Code or adapted for specific needs. The agent template system provides a wizard for creating specialized sub-agents (e.g., security auditors, systems architects) with standardized formatting and proper tool access. The repository is under active development, so caution is advised when using it in production environments.

AgentFly
AgentFly is an extensible framework for building LLM agents with reinforcement learning. It supports multi-turn training by adapting traditional RL methods with token-level masking. It features a decorator-based interface for defining tools and reward functions, enabling seamless extension and ease of use. To support high-throughput training, it implemented asynchronous execution of tool calls and reward computations, and designed a centralized resource management system for scalable environment coordination. A suite of prebuilt tools and environments are provided.

SeerAttention
SeerAttention is a novel trainable sparse attention mechanism that learns intrinsic sparsity patterns directly from LLMs through self-distillation at post-training time. It achieves faster inference while maintaining accuracy for long-context prefilling. The tool offers features such as trainable sparse attention, block-level sparsity, self-distillation, efficient kernel, and easy integration with existing transformer architectures. Users can quickly start using SeerAttention for inference with AttnGate Adapter and training attention gates with self-distillation. The tool provides efficient evaluation methods and encourages contributions from the community.

TheoremExplainAgent
TheoremExplainAgent is an AI system that generates long-form Manim videos to visually explain theorems, proving its deep understanding while uncovering reasoning flaws that text alone often hides. The codebase for the paper 'TheoremExplainAgent: Towards Multimodal Explanations for LLM Theorem Understanding' is available in this repository. It provides a tool for creating multimodal explanations for theorem understanding using AI technology.

starcoder2-self-align
StarCoder2-Instruct is an open-source pipeline that introduces StarCoder2-15B-Instruct-v0.1, a self-aligned code Large Language Model (LLM) trained with a fully permissive and transparent pipeline. It generates instruction-response pairs to fine-tune StarCoder-15B without human annotations or data from proprietary LLMs. The tool is primarily finetuned for Python code generation tasks that can be verified through execution, with potential biases and limitations. Users can provide response prefixes or one-shot examples to guide the model's output. The model may have limitations with other programming languages and out-of-domain coding tasks.

crab
CRAB is a framework for building LLM agent benchmark environments in a Python-centric way. It is cross-platform and multi-environment, allowing the creation of agent environments supporting various deployment options. The framework offers easy-to-use configuration with the ability to add new actions and define environments seamlessly. CRAB also provides a novel benchmarking suite with tasks and evaluators defined in Python, along with a unique graph evaluator method for detailed metrics.

PowerInfer
PowerInfer is a high-speed Large Language Model (LLM) inference engine designed for local deployment on consumer-grade hardware, leveraging activation locality to optimize efficiency. It features a locality-centric design, hybrid CPU/GPU utilization, easy integration with popular ReLU-sparse models, and support for various platforms. PowerInfer achieves high speed with lower resource demands and is flexible for easy deployment and compatibility with existing models like Falcon-40B, Llama2 family, ProSparse Llama2 family, and Bamboo-7B.

data-juicer
Data-Juicer is a one-stop data processing system to make data higher-quality, juicier, and more digestible for LLMs. It is a systematic & reusable library of 80+ core OPs, 20+ reusable config recipes, and 20+ feature-rich dedicated toolkits, designed to function independently of specific LLM datasets and processing pipelines. Data-Juicer allows detailed data analyses with an automated report generation feature for a deeper understanding of your dataset. Coupled with multi-dimension automatic evaluation capabilities, it supports a timely feedback loop at multiple stages in the LLM development process. Data-Juicer offers tens of pre-built data processing recipes for pre-training, fine-tuning, en, zh, and more scenarios. It provides a speedy data processing pipeline requiring less memory and CPU usage, optimized for maximum productivity. Data-Juicer is flexible & extensible, accommodating most types of data formats and allowing flexible combinations of OPs. It is designed for simplicity, with comprehensive documentation, easy start guides and demo configs, and intuitive configuration with simple adding/removing OPs from existing configs.

aircrack-ng
Aircrack-ng is a comprehensive suite of tools designed to evaluate the security of WiFi networks. It covers various aspects of WiFi security, including monitoring, attacking (replay attacks, deauthentication, fake access points), testing WiFi cards and driver capabilities, and cracking WEP and WPA PSK. The tools are command line-based, allowing for extensive scripting and have been utilized by many GUIs. Aircrack-ng primarily works on Linux but also supports Windows, macOS, FreeBSD, OpenBSD, NetBSD, Solaris, and eComStation 2.

smile
Smile (Statistical Machine Intelligence and Learning Engine) is a comprehensive machine learning, NLP, linear algebra, graph, interpolation, and visualization system in Java and Scala. It covers every aspect of machine learning, including classification, regression, clustering, association rule mining, feature selection, manifold learning, multidimensional scaling, genetic algorithms, missing value imputation, efficient nearest neighbor search, etc. Smile implements major machine learning algorithms and provides interactive shells for Java, Scala, and Kotlin. It supports model serialization, data visualization using SmilePlot and declarative approach, and offers a gallery showcasing various algorithms and visualizations.

CuMo
CuMo is a project focused on scaling multimodal Large Language Models (LLMs) with Co-Upcycled Mixture-of-Experts. It introduces CuMo, which incorporates Co-upcycled Top-K sparsely-gated Mixture-of-experts blocks into the vision encoder and the MLP connector, enhancing the capabilities of multimodal LLMs. The project adopts a three-stage training approach with auxiliary losses to stabilize the training process and maintain a balanced loading of experts. CuMo achieves comparable performance to other state-of-the-art multimodal LLMs on various Visual Question Answering (VQA) and visual-instruction-following benchmarks.
For similar tasks

understand-r1-zero
The 'understand-r1-zero' repository focuses on understanding R1-Zero-like training from a critical perspective. It provides insights into base models and reinforcement learning components, highlighting findings and proposing solutions for biased optimization. The repository offers a minimalist recipe for R1-Zero training, detailing the RL-tuning process and achieving state-of-the-art performance with minimal compute resources. It includes codebase, models, and paper related to R1-Zero training implemented with the Oat framework, emphasizing research-friendly and efficient LLM RL techniques.

lighteval
LightEval is a lightweight LLM evaluation suite that Hugging Face has been using internally with the recently released LLM data processing library datatrove and LLM training library nanotron. We're releasing it with the community in the spirit of building in the open. Note that it is still very much early so don't expect 100% stability ^^' In case of problems or question, feel free to open an issue!

Firefly
Firefly is an open-source large model training project that supports pre-training, fine-tuning, and DPO of mainstream large models. It includes models like Llama3, Gemma, Qwen1.5, MiniCPM, Llama, InternLM, Baichuan, ChatGLM, Yi, Deepseek, Qwen, Orion, Ziya, Xverse, Mistral, Mixtral-8x7B, Zephyr, Vicuna, Bloom, etc. The project supports full-parameter training, LoRA, QLoRA efficient training, and various tasks such as pre-training, SFT, and DPO. Suitable for users with limited training resources, QLoRA is recommended for fine-tuning instructions. The project has achieved good results on the Open LLM Leaderboard with QLoRA training process validation. The latest version has significant updates and adaptations for different chat model templates.

Awesome-Text2SQL
Awesome Text2SQL is a curated repository containing tutorials and resources for Large Language Models, Text2SQL, Text2DSL, Text2API, Text2Vis, and more. It provides guidelines on converting natural language questions into structured SQL queries, with a focus on NL2SQL. The repository includes information on various models, datasets, evaluation metrics, fine-tuning methods, libraries, and practice projects related to Text2SQL. It serves as a comprehensive resource for individuals interested in working with Text2SQL and related technologies.

create-million-parameter-llm-from-scratch
The 'create-million-parameter-llm-from-scratch' repository provides a detailed guide on creating a Large Language Model (LLM) with 2.3 million parameters from scratch. The blog replicates the LLaMA approach, incorporating concepts like RMSNorm for pre-normalization, SwiGLU activation function, and Rotary Embeddings. The model is trained on a basic dataset to demonstrate the ease of creating a million-parameter LLM without the need for a high-end GPU.

StableToolBench
StableToolBench is a new benchmark developed to address the instability of Tool Learning benchmarks. It aims to balance stability and reality by introducing features such as a Virtual API System with caching and API simulators, a new set of solvable queries determined by LLMs, and a Stable Evaluation System using GPT-4. The Virtual API Server can be set up either by building from source or using a prebuilt Docker image. Users can test the server using provided scripts and evaluate models with Solvable Pass Rate and Solvable Win Rate metrics. The tool also includes model experiments results comparing different models' performance.

BetaML.jl
The Beta Machine Learning Toolkit is a package containing various algorithms and utilities for implementing machine learning workflows in multiple languages, including Julia, Python, and R. It offers a range of supervised and unsupervised models, data transformers, and assessment tools. The models are implemented entirely in Julia and are not wrappers for third-party models. Users can easily contribute new models or request implementations. The focus is on user-friendliness rather than computational efficiency, making it suitable for educational and research purposes.

AI-TOD
AI-TOD is a dataset for tiny object detection in aerial images, containing 700,621 object instances across 28,036 images. Objects in AI-TOD are smaller with a mean size of 12.8 pixels compared to other aerial image datasets. To use AI-TOD, download xView training set and AI-TOD_wo_xview, then generate the complete dataset using the provided synthesis tool. The dataset is publicly available for academic and research purposes under CC BY-NC-SA 4.0 license.
For similar jobs

Perplexica
Perplexica is an open-source AI-powered search engine that utilizes advanced machine learning algorithms to provide clear answers with sources cited. It offers various modes like Copilot Mode, Normal Mode, and Focus Modes for specific types of questions. Perplexica ensures up-to-date information by using SearxNG metasearch engine. It also features image and video search capabilities and upcoming features include finalizing Copilot Mode and adding Discover and History Saving features.

KULLM
KULLM (ꡬλ¦) is a Korean Large Language Model developed by Korea University NLP & AI Lab and HIAI Research Institute. It is based on the upstage/SOLAR-10.7B-v1.0 model and has been fine-tuned for instruction. The model has been trained on 8ΓA100 GPUs and is capable of generating responses in Korean language. KULLM exhibits hallucination and repetition phenomena due to its decoding strategy. Users should be cautious as the model may produce inaccurate or harmful results. Performance may vary in benchmarks without a fixed system prompt.

MMMU
MMMU is a benchmark designed to evaluate multimodal models on college-level subject knowledge tasks, covering 30 subjects and 183 subfields with 11.5K questions. It focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. The evaluation of various models highlights substantial challenges, with room for improvement to stimulate the community towards expert artificial general intelligence (AGI).

1filellm
1filellm is a command-line data aggregation tool designed for LLM ingestion. It aggregates and preprocesses data from various sources into a single text file, facilitating the creation of information-dense prompts for large language models. The tool supports automatic source type detection, handling of multiple file formats, web crawling functionality, integration with Sci-Hub for research paper downloads, text preprocessing, and token count reporting. Users can input local files, directories, GitHub repositories, pull requests, issues, ArXiv papers, YouTube transcripts, web pages, Sci-Hub papers via DOI or PMID. The tool provides uncompressed and compressed text outputs, with the uncompressed text automatically copied to the clipboard for easy pasting into LLMs.

gpt-researcher
GPT Researcher is an autonomous agent designed for comprehensive online research on a variety of tasks. It can produce detailed, factual, and unbiased research reports with customization options. The tool addresses issues of speed, determinism, and reliability by leveraging parallelized agent work. The main idea involves running 'planner' and 'execution' agents to generate research questions, seek related information, and create research reports. GPT Researcher optimizes costs and completes tasks in around 3 minutes. Features include generating long research reports, aggregating web sources, an easy-to-use web interface, scraping web sources, and exporting reports to various formats.

ChatTTS
ChatTTS is a generative speech model optimized for dialogue scenarios, providing natural and expressive speech synthesis with fine-grained control over prosodic features. It supports multiple speakers and surpasses most open-source TTS models in terms of prosody. The model is trained with 100,000+ hours of Chinese and English audio data, and the open-source version on HuggingFace is a 40,000-hour pre-trained model without SFT. The roadmap includes open-sourcing additional features like VQ encoder, multi-emotion control, and streaming audio generation. The tool is intended for academic and research use only, with precautions taken to limit potential misuse.

HebTTS
HebTTS is a language modeling approach to diacritic-free Hebrew text-to-speech (TTS) system. It addresses the challenge of accurately mapping text to speech in Hebrew by proposing a language model that operates on discrete speech representations and is conditioned on a word-piece tokenizer. The system is optimized using weakly supervised recordings and outperforms diacritic-based Hebrew TTS systems in terms of content preservation and naturalness of generated speech.

do-research-in-AI
This repository is a collection of research lectures and experience sharing posts from frontline researchers in the field of AI. It aims to help individuals upgrade their research skills and knowledge through insightful talks and experiences shared by experts. The content covers various topics such as evaluating research papers, choosing research directions, research methodologies, and tips for writing high-quality scientific papers. The repository also includes discussions on academic career paths, research ethics, and the emotional aspects of research work. Overall, it serves as a valuable resource for individuals interested in advancing their research capabilities in the field of AI.