OpenR1-Qwen-7B

This is a finetune of Qwen2.5-Math-Instruct on OpenR1-220k-Math (default split).

Check out OpenR1-Distill-7B for an improved model that was trained on open-r1/Mixture-of-Thoughts and replicates the performance of DeepSeek-R1-Distill-Qwen-7B across multiple reasoning domains.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "open-r1/OpenR1-Qwen-7B"
device = "cuda" 

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$."

messages = [
    {"role": "system", "content": "Please reason step by step, and put your final answer within \\boxed{}."},
    {"role": "user", "content": prompt}
]

Training

We train the model on the default split of OpenR1-220k-Math for 3 epochs. We use learning rate of 5e-5 and extend the context length from 4k to 32k, by increasing RoPE frequency to 300k. The training follows a linear learning rate schedule with a 10% warmup phase. The table below compares the performance of OpenR1-Qwen-7B to DeepSeek-R1-Distill-Qwen-7B and OpenThinker-7B using lighteval.

You can find the training and evaluation code at: https://github.com/huggingface/open-r1/

Model	MATH-500	AIME 2024	AIME 2025	GPQA-D
DeepSeek-Distill-Qwen-7B	93.5	51.3	35.8	52.4
OpenR1-Qwen-7B	90.6	47.0	33.2	42.4
OpenThinker-7B	86.4	31.3	24.6	39.1