---
language: en
license: apache-2.0
library_name: transformers
pipeline_tag: text-classification
tags:
- router
- reasoning
- text-classification
- hybrid
- qwen3
model_name: Reasoning Router 0.6B
datasets:
- AmirMohseni/reasoning-router-data-v2
base_model:
- Qwen/Qwen3-0.6B
---

# Reasoning Router 0.6B

**AmirMohseni/reasoning-router-0.6b** is a fine-tuned reasoning router built on top of `Qwen/Qwen3-0.6B`. It classifies user prompts into two categories:
* `no_think` → The task does not require explicit reasoning.
* `think` → The task benefits from a reasoning mode (e.g., math, multi-step analysis).

This router is designed for hybrid model systems, where it decides whether to route prompts to lightweight inference endpoints or to reasoning-enabled models such as the Qwen3 series or `deepseek-ai/DeepSeek-V3.1`.

---

## Use Case

The reasoning router allows for efficient orchestration in model pipelines:
* Run cheap, fast inference for simple tasks.
* Switch to more powerful, expensive reasoning models only when needed.

This approach helps **reduce costs, latency, and unnecessary compute** in real-world deployments.

---

## 🚀 Quick Start

### Example Usage

```python
from transformers import pipeline

# Initialize the router pipeline
router = pipeline(
    "text-classification",
    model="AmirMohseni/reasoning-router-0.6b",
    device_map="auto"
)

# Example prompt that requires reasoning
prompt = "What is the sum of the first 100 prime numbers?"
results = router(prompt)[0]

print('Label: ', results['label']) # Label:  no_think
print('Probability Score: ', results['score']) # Probability Score:  0.6192409992218018
```

-----

## 📚 Training Data

This model was trained on the `AmirMohseni/reasoning-router-data-v2` dataset, which was curated from multiple instruction-following datasets. The dataset primarily contains:

  * **Math reasoning data** → Derived from Big-Math-RL and AIME problems (1983–2024).
  * **General tasks** → A mix of simple vs. reasoning-heavy queries to teach the model to distinguish between them.

-----

## ⚠️ Limitations

  * **Language Coverage**: The model is trained primarily on English. Performance on other languages may be weaker.
  * **Reasoning Coverage**: For tasks labeled `think`, the training data is heavily skewed towards mathematical reasoning.
  * **No Coding Tasks**: Programming or code-related reasoning tasks are not included in the current training data.

-----

## 🔧 Model Details

  * **Base model**: `Qwen/Qwen3-0.6B`
  * **Parameters**: 0.6B
  * **Task**: Binary classification (`no_think`, `think`)
  * **Intended use**: Routing prompts for hybrid reasoning pipelines.

-----

## ✅ Intended Use

  * Routing user prompts in a multi-model reasoning system.
  * Reducing compute costs by filtering out tasks that don’t require a dedicated reasoning model.