max_steps: 450
per_device_train_batch_size: 1
gradient_accumulation_steps: 96
generation_batch_size: 32
gradient_checkpointing: true
gradient_checkpointing_kwargs:
- use_reentrant: false
learning_rate: 5.0e-7 # 1.0e-6 as in the deepseek math paper 5-e7 from https://hijkzzz.notion.site/unraveling-rlhf-and-its-variants-engineering-insights#147d9a33ecc9806090f3d5c749d31f05
lr_scheduler_type: cosine
warmup_ratio: 0.03

GRPO specific parameters

beta: 0.001 # 0.04 as in the deepseek math paper 0.001 from https://hijkzzz.notion.site/unraveling-rlhf-and-its-variants-engineering-insights#147d9a33ecc9806090f3d5c749d31f05
max_prompt_length: 256
max_completion_length: 1024
num_generations: 8
use_vllm: true
vllm_mode: colocate
vllm_gpu_memory_utilization: 0.2

This model is a fine-tuned version of Qwen/Qwen2.5-3B-Instruct. It has been trained using TRL.

Quick start

python

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="vkasera/v3_qwen-2.5-3b-r1-countdown-phil", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.

Framework versions

TRL: 0.23.1
Transformers: 4.56.2
Pytorch: 2.7.0
Datasets: 4.1.1
Tokenizers: 0.22.1

Citations

Cite GRPO as:

bibtex

@article{shao2024deepseekmath,
    title        = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
    author       = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
    year         = 2024,
    eprint       = {arXiv:2402.03300},
}

Cite TRL as:

bibtex

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}