EvolLLM/README.md

---
base_model:
- Qwen/Qwen3-4B-Instruct-2507
- Qwen/Qwen3-4B-Thinking-2507
library_name: transformers
datasets:
- openai/gsm8k
tags:
- evoluation
- math
- merge
---

# 📑 Model Card


[![GitHub](https://img.shields.io/badge/GitHub-181717?style=flat-square&logo=github&logoColor=white)](https://github.com/Hert4)
[![HuggingFace](https://img.shields.io/badge/HuggingFace-FFD21E?style=flat-square&logo=huggingface&logoColor=black)](https://huggingface.co/beyoru)
[![BMC](https://img.shields.io/badge/buy_me_a_coffee-A78BFA?style=flat-square&logo=buy-me-a-coffee&logoColor=white)](https://buymeacoffee.com/ductransa0g)

## Model Details
This model is a merged version of two Qwen base models:  
- **Qwen/Qwen3-4B-Instruct-2507**  
- **Qwen/Qwen3-4B-Thinking-2507**  


## Notations:
- **Evoluation dataset**: `openai/gsm8k` (subset of 100 samples, not trained)  
- **Generation runs**: 50  
- **Population size**: 10
- This model design for instruct model not reasoning model with same function like Qwen3-Instruct-2507
- **A good start for SFT or GRPO training.**

## Evaluation

- For my evaluation in my agent benchmark is not surpass too much but only 3% with instruct model.
- Surpass `openfree/Darwin-Qwen3-4B` (Evolution model) and base model in ACEBench.


```bibtex
@misc{nafy_qwen_merge_2025,
  title        = {Merged Qwen3 4B Instruct + Thinking Models},
  author       = {Beyoru},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/beyoru/EvolLLM}},
  note         = {Merged model combining instruction-tuned and reasoning Qwen3 variants.},
  base_models  = {Qwen/Qwen3-4B-Instruct-2507, Qwen/Qwen3-4B-Thinking-2507}
}