🧬 Evol-Aes-Hybrid-4B: Evolutionary Optimized Merge

💻 GitHub Repo • 🤗 Model Collection

📖 Overview

Evol-Aes-Hybrid-4B is a high-performance 4B parameter model created using an advanced Evolutionary Layer-Wise Merge strategy. By combining the generalist capabilities of Instruct with the logical prowess of Thinking models, this hybrid achieves a superior balance between conversational fluency and reasoning depth.

This model is designed to be a robust starting point for SFT (Supervised Fine-Tuning) or GRPO training, offering a stabilized foundation that outperforms standard merges.

🧩 Base Models

This model is a genetic evolution of the following parents:

• Base (Generalist): Qwen/Qwen3-4B-Instruct-2507
• Source (Specialist): Qwen/Qwen3-4B-Thinking-2507

💡 Design Philosophy: Pure thinking models often over-reason on simple tasks, while instruct models may lack depth. This hybrid applies a "genetic selection" to retain the best layers from both worlds.

---

⚙️ Evolutionary Methodology

Unlike traditional linear merges (SLERP/TIES), this model was optimized using a Genetic Algorithm (GA) to find the perfect weight ratio for each individual layer.

🔬 Experiment Settings

• Optimization Target: Minimized metrics on the openai/gsm8k dataset.
• Evaluation Subset: 100 challenging samples.
• Population Size: 10 candidates per generation.
• Generations: 30 evolution cycles.
• Algorithm: Layer-wise weighted averaging with mutation and crossover (Elitism strategy).

🧬 Layer Composition Analysis

The evolutionary process discovered a non-linear structure:

1. Early Layers (0-5): Heavily favors the Instruct Base (~65%) to maintain robust language understanding and grammar.
2. Middle Layers: Incorporates Thinking modules for logic processing.
3. Specific Heads: Selectively rejected unstable layers from the Thinking model (e.g., layers with high Alpha metrics) to ensure stability.

---

📊 Evaluation & Performance

Rank	Model	Overall Score
1	Evol-Aes-Hybrid-4B	0.699
2	EvoLLM	0.663
3	Qwen3-2507 (Base)	0.639
4	Darwin-Qwen3-4B	0.610

---

💻 Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "beyoru/Evol-Aes-Hybrid-4B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Solve 2x+1=3 !!!"},
]

input_ids = tokenizer.apply_chat_template(
    messages, 
    add_generation_prompt=True, 
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=512,
)

print(tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True))

🤗 Support me at:

---

📜 Citation

If you use this model or the evolutionary merge method in your work, please cite:

@misc{nafy_qwen_merge_2025,
  title        = {EvolLLM: Evolutionary Optimized Qwen3 Hybrid},
  author       = {Beyoru},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{[https://huggingface.co/beyoru/Evol-Aes-Hybrid-4B](https://huggingface.co/beyoru/Evol-Aes-Hybrid-4B)}},
  note         = {Merged model combining instruction-tuned and reasoning Qwen3 variants via Genetic Algorithm.},
  base_models  = {Qwen/Qwen3-4B-Instruct-2507, Qwen/Qwen3-4B-Thinking-2507}
}