RLVR-SvS/SvS-Qwen-Code-7B
Model Card for SvS-Code-7B (from Qwen2.5-7B-Instruct)
<p align="left"> <a href="https://mastervito.github.io/SvS.github.io/"><b>[🌐 Website]</b></a> • <a href="https://huggingface.co/datasets/RLVR-SvS/Variational-DAPO"><b>[🤗 Dataset]</b></a> • <a href="https://huggingface.co/RLVR-SvS/SvS-Qwen-32B"><b>[🤖 Models]</b></a> • <a href="https://arxiv.org/abs/2508.14029"><b>[📜 Paper]</b></a> • <a href="https://github.com/MasterVito/SvS"><b>[🐱 GitHub]</b></a> • <a href="https://huggingface.co/datasets/RLVR-SvS/Variational-DAPO"><b>[🐦 Twitter]</b></a> • <a href="https://huggingface.co/datasets/RLVR-SvS/Variational-DAPO"><b>[📕 Rednote]</b></a> </p>
The official model checkpoints for <a href="https://arxiv.org/abs/2508.14029"><b>SvS</b></a>. The SvS model is trained on a subset of coding tasks from PRIME-RL dataset (included in this repository as <code>12kcoderl.parquet</code>).
Inference
We recommend using our official inference template from Qwen2.5 Instruct models.
model_name = "RLVR-SvS/SvS-Qwen-Code-7B"
device = "cuda" # the device to load the model onto
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "write a quick sort algorithm."
messages = [
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
generated_ids = model.generate(
**model_inputs,
max_new_tokens=8192
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]Cite Us
If you find the model helpful, please consider citing our paper:
@misc{liang2025pass1selfplayvariationalproblem,
title={Beyond Pass@1: Self-Play with Variational Problem Synthesis Sustains RLVR},
author={Xiao Liang and Zhongzhi Li and Yeyun Gong and Yelong Shen and Ying Nian Wu and Zhijiang Guo and Weizhu Chen},
year={2025},
eprint={2508.14029},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2508.14029},
}