Qwen2.5-7B-MathSoup

🍲 Model Soup created using weighted averaging based on Meta's Souper-Model.

Weights

• math: 60%
• general: 40%

Expected Performance (Linear Prediction)

Benchmark	Predicted Score
GSM8K	88.3%
HumanEval	59.5%

Note: Actual performance may differ due to weight interference effects.

Component Models

Model	GSM8K	HumanEval
Qwen2.5-7B-Instruct	85.4%	70.1%
Qwen2.5-Coder-7B-Instruct	60.4%	88.4%
Qwen2.5-Math-7B-Instruct	90.3%	52.4%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("researchaudio/Qwen2.5-7B-MathSoup")
tokenizer = AutoTokenizer.from_pretrained("researchaudio/Qwen2.5-7B-MathSoup")

messages = [{"role": "user", "content": "Solve: What is 15% of 80?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0]))

Citation

@misc{soupermodel2025,
    title={Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance},
    author={Shalini Maiti and others},
    year={2025},
    url={https://arxiv.org/abs/2511.13254},
}