---
base_model:
- Menlo/Jan-nano
- POLARIS-Project/Polaris-4B-Preview
- Qwen/Qwen3-4B
library_name: transformers
tags:
- mergekit
- merge

---
# Qwen3-4B-Agentic-Reasoner

  <img src="modelcard.png" widhth="400"/>

`yasserrmd/qwen3-4b-agentic-reasoner` is a merged model that combines the **agentic instruction-following strength** of [Menlo/Jan-nano](https://huggingface.co/Menlo/Jan-nano) with the **reasoning and structured thought capabilities** of [POLARIS-Project/Polaris-4B-Preview](https://huggingface.co/POLARIS-Project/Polaris-4B-Preview), using the [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) architecture as the base.

This merge was performed using [mergekit](https://github.com/cg123/mergekit) and the [TIES](https://arxiv.org/abs/2306.01708) method for fine-grained parameter blending.

---

## 🧠 Intended Use

This model is intended for use in:

- Multi-step reasoning tasks  
- Agent-style instruction following (CLI assistants, web automation)  
- Educational assistance, planning, and explanation  
- Natural language code generation, JSON/schema design  
- Legal, productivity, and roleplay simulations

---

## 🧪 Merge Details

### 🔀 Merge Method

This model was merged using the [TIES](https://arxiv.org/abs/2306.01708) merge method with the [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) as the base model.

### 🤝 Models Merged

| Model | Role |
|-------|------|
| [POLARIS-Project/Polaris-4B-Preview](https://huggingface.co/POLARIS-Project/Polaris-4B-Preview) | Deep reasoning & CoT |
| [Menlo/Jan-nano](https://huggingface.co/Menlo/Jan-nano) | Agentic & Instruction-following |

### ⚙️ Configuration

```yaml
models:
  - model: POLARIS-Project/Polaris-4B-Preview
    parameters:
      weight: 0.5
  - model: Menlo/Jan-nano
    parameters:
      weight: 0.5
merge_method: ties
base_model: Qwen/Qwen3-4B
parameters:
  normalize: true
  int8_mask: true
dtype: float16
````

---

## 📊 Prompt Evaluation

This model was evaluated on **handcrafted prompts** covering:

* Chain-of-thought reasoning
* Math and logic
* Code writing and CLI instructions
* JSON/schema generation
* Role-based planning and writing tasks
* Arabic translation
* Legal drafting

### ✅ Performance Highlights

| Criterion                  | Result                                  |
| -------------------------- | --------------------------------------- |
| CoT Reasoning              | Excellent (multi-step math, planning)   |
| Agentic Tasks              | Strong (shell scripts, terminal agents) |
| Code Output                | Clean formatting and logical structure  |
| Format Awareness           | Recognizes JSON, email, legal structure |
| Instruction Follow-through | Reliable and contextual                 |
| Language Tasks             | Accurate Arabic translation, paraphrase |

Average prompt score (0–3 scale): **2.15**
All outputs were logical, well-structured, and contextually accurate for the prompt types.

---

## 🚀 Inference

To use the model:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "yasserrmd/qwen3-4b-agentic-reasoner"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto", trust_remote_code=True)

prompt = "Plan the first 3 steps for launching a nonprofit AI education platform."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## ⚠️ License & Use

Respect the licenses of the original merged models. This model is released for **research and personal experimentation** purposes.

* [POLARIS-Project/Polaris-4B-Preview License](https://huggingface.co/POLARIS-Project/Polaris-4B-Preview)
* [Menlo/Jan-nano License](https://huggingface.co/Menlo/Jan-nano)
* [Qwen3-4B License](https://huggingface.co/Qwen/Qwen3-4B)

---

## 🙏 Acknowledgments

Thanks to the teams behind:

* Alibaba's Qwen3 series
* Menlo/Jan-nano project
* POLARIS RL framework
* MergeKit by [@cg123](https://github.com/cg123)

---

Model by [@yasserrmd](https://huggingface.co/yasserrmd)