This model is a fine-tuned version of
deepseek-ai/DeepSeek-R1-Distill-Qwen-32B on the
S1.1-QwQ dataset.
The model has achieved state-of-the-art reasoning capabilities on challengining benchmarks including AIME2024/2025, MATH500 and GPQA-Diamond.
We utilize
LLaMAFactory with $8\times A100-SXM4-80GB$ GPU to conduct full-parameter finetuning on our self-curated S1.1-QWQ dataset, which is another refined version of
S1.1-1K dataset.
We use
QwQ-32B to generate reasoning trajectories for each of the problem in S1.1-1k dataset. The experiment turns out that the quality of QwQ generated trajectories are better than the original version including (Gemini-2.0-flash-thinking and DeepSeek-R1).
Here we present the evaluation results of our S1.1-QwQ-DS/Qwen-32B on challenging reasoning tasks including AIME2024,AIM2025,MATH500 and GPQA-Diamond.
Compare to other version of s1-1k dataset, our newly curated dataset has demonstrate the supeority of performance gains based on
Qwen2.5-32B-Instruct and
DeepSeek-R1-Distill-Qwen-32B over all benchmarks.
1from vllm import LLM, SamplingParams
2from transformers import AutoTokenizer
3model_id = 'BitStarWalkin/S1.1-QwQ-DS'
4model = LLM(
5 model_id,
6 tensor_parallel_size=8,
7)
8tokenizer = AutoTokenizer.from_pretrained(model_id)
9sampling_params = SamplingParams(
10 max_tokens=16384,
11)
12question = """Let \(x, y\), and \(z\) be positive real numbers satisfying the system of equations:
13\[
14\begin{array}{c}
15\sqrt{2 x-x y}+\sqrt{2 y-x y}=1 \\
16\sqrt{2 y-y z}+\sqrt{2 z-y z}=\sqrt{2} \\
17\sqrt{2 z-z x}+\sqrt{2 x-z x}=\sqrt{3} .
18\end{array}
19\]
20Then \(\left[(1-x)(1-y)(1-z)\right]^{2}\) can be written as \(\frac{m}{n}\), where \(m\) and \(n\) are relatively prime positive integers. Find \(m+n\)."""
21ds_prompt="<｜User｜>\n" + question + "<｜Assistant｜>\n"
22output = model.generate(ds_prompt, sampling_params=sampling_params)
23print(output[0].outputs[0].text)
