README.md

---
license: apache-2.0
language:
- en
tags:
- reasoning
- chain-of-thought
- qwen
- tiny
- whirlwindai
pipeline_tag: text-generation
datasets:
- WhirlwindAI/Soft-CoT-1K
library_name: transformers
base_model:
- Qwen/Qwen2.5-0.5B-Instruct
---

<div align="center">

<img src="https://capsule-render.vercel.app/api?type=waving&height=220&color=gradient&customColorList=12,19,24,30&text=Qwen-R1-0.5B&fontSize=48&fontColor=ffffff&animation=twinkling"/>

<img src="https://readme-typing-svg.demolab.com?font=Space+Grotesk&weight=700&size=27&duration=2300&pause=1200&color=A855F7&center=true&vCenter=true&width=850&lines=Qwen-R1-0.5B;Reason+First.+Answer+Second.;Chain-of-Thought+on+a+Tiny+Model." />

<img src="https://img.shields.io/badge/Parameters-0.5B-A855F7?style=for-the-badge">
<img src="https://img.shields.io/badge/Base-Qwen2.5--0.5B-7C3AED?style=for-the-badge">
<img src="https://img.shields.io/badge/Trained%20On-Soft--CoT--1K-06B6D4?style=for-the-badge">
<img src="https://img.shields.io/badge/License-Apache--2.0-22C55E?style=for-the-badge">

</div>

---

# 💡 The Idea

<div align="center">

> **Good answers come from good thinking.**

Qwen-R1-0.5B is a fine-tuned version of Qwen2.5-0.5B-Instruct trained to **reason before it answers** using explicit `<thinking>` tags.

</div>

Instead of jumping straight to the answer, this model generates its reasoning first — making it more transparent, more reliable, and easier to debug.

---

# 🧠 How It Works

Every response is structured as:

```
User: {question}
Assistant: <thinking>{reasoning}</thinking>
{answer}
```

The model learns to:
1. **Think** – generate step-by-step reasoning
2. **Answer** – provide the final response

---

# 📊 Training Details

| Property | Value |
|----------|-------|
| Base Model | Qwen2.5-0.5B-Instruct |
| Dataset | WhirlwindAI/Soft-CoT-1K |
| Examples | 1,355 |
| Method | QLoRA (4-bit) |
| Epochs | 3 |
| Learning Rate | 2e-4 |
| LoRA Rank | 16 |
| LoRA Alpha | 32 |

---

# 🚀 Quick Start

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "WhirlwindAI/Qwen-R1-0.5B"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

prompt = "User: What is 2+2?\nAssistant: <thinking>"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50, do_sample=True, temperature=0.7)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

# 📋 Sample Output

```
User: What is the capital of France?
Assistant: <thinking>Paris is the capital of France.</thinking>
Paris
```

---

# 📈 Performance

The model was evaluated on 10 out-of-distribution questions:

| Category | Performance |
|----------|-------------|
| Format (thinking tags) | ✅ Excellent |
| General Knowledge | ✅ Good |
| Creative Reasoning | ✅ Good |
| Math/Logic | ⚠️ Needs improvement |
| Physics/Science | ⚠️ Needs improvement |

---

# 🔬 What It Learned

| Strength | Weakness |
|----------|----------|
| ✅ Consistent `<thinking>` format | ❌ Sometimes hallucinates facts |
| ✅ Generates reasoning before answering | ❌ Struggles with multi-step math |
| ✅ Retains general knowledge | ❌ Physics reasoning needs more data |

---

# 🧪 Test It Yourself

```python
questions = [
    "What is the capital of France?",
    "Explain entropy like I'm 5.",
    "Write a short poem about a robot.",
]

for q in questions:
    prompt = f"User: {q}\nAssistant: <thinking>"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=80)
    print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

# 📜 Citation

```bibtex
@model{qwenr1_2026,
  title={Qwen-R1-0.5B},
  author={WhirlwindAI},
  year={2026},
  publisher={Hugging Face}
}
```

---

<div align="center">

### 🌪️ WhirlwindAI

**Efficient Models • Practical Research • Open AI**

<br>

<img src="https://capsule-render.vercel.app/api?type=waving&height=140&section=footer&color=0:A855F7,100:06B6D4"/>

</div>