TinyZero: A Low-Cost Replication of DeepSeeK-R1 Zero's Epiphany Effect

General Introduction

TinyZero is a veRL-based reinforcement learning model designed to reproduce the DeepSeeK-R1 Zero's performance in countdown and multiplication tasks. Surprisingly, the program was able to achieve the same epiphanies as DeepSeeK-R1 Zero for a running cost of only $30 (less than 5 hours using 2xH200 at $6.4 per hour). Through Reinforcement Learning (RL), the 3B Base Language Model (LM) is able to autonomously develop self-validation and search capabilities. Users can experience the power and innovation of TinyZero through a simple setup and training process.

Function List

• countdown task: Supports data preparation and training processes to help models learn in countdown tasks.
• Multiplication tasks: Supports data preparation and training processes to help models learn in multiplication tasks.
• Single GPU Support: For model parameters less than or equal to 1.5B.
• Multi-GPU Support: Models applicable to larger parameters are capable of developing sophisticated reasoning.
• Instruct Ablation: Experiments supporting the QWen-2.5-3B Instruct model.
• Quality Improvement ToolsThe tools include flash-attn, wandb, IPython, and matplotlib to enhance the model training and usage experience.

Using Help

Installation process

1. Create a virtual environment:

   conda create -n zero python=3.9
   

2. Install PyTorch (optional):

   pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu121
   

3. Install vllm:

   pip3 install vllm==0.6.3
   

4. Install ray:

   pip3 install ray
   

5. Install verl:

   pip install -e .
   

6. Install flash-attn:

   pip3 install flash-attn --no-build-isolation
   

7. Installation of quality enhancement tools:

   pip install wandb IPython matplotlib
   

Functional operation flow

countdown task

1. Data preparation:

   conda activate zero
   python . /examples/data_preprocess/countdown.py --local_dir {path_to_your_dataset}
   

2. Training process:

   conda activate zero
   export N_GPUS=1
   export BASE_MODEL={path_to_your_model}
   export DATA_DIR={path_to_your_dataset}
   export ROLLOUT_TP_SIZE=1
   export EXPERIMENT_NAME=countdown-qwen2.5-0.5b
   export VLLM_ATTENTION_BACKEND=XFORMERS
   bash . /scripts/train_tiny_zero.sh
   

3B+ Model Training

1. Data preparation:

   conda activate zero
   python examples/data_preprocess/countdown.py --template_type=qwen-instruct --local_dir={path_to_your_dataset}
   

2. Training process:

   conda activate zero
   export N_GPUS=2
   export BASE_MODEL={path_to_your_model}
   export DATA_DIR={path_to_your_dataset}
   export ROLLOUT_TP_SIZE=2
   export EXPERIMENT_NAME=countdown-qwen2.5-3b-instruct
   export VLLM_ATTENTION_BACKEND=XFORMERS
   bash . /scripts/train_tiny_zero.sh