---
library_name: transformers
tags:
- unsloth
- trl
- grpo
license: mit
datasets:
- eagle0504/openai-gsm8k-enhanced-using-together-ai-deepseek-train8k-test1k-v1
language:
- en
base_model:
- Qwen/Qwen2.5-1.5B-Instruct
---

# Qwen2.5-1.5B-Instruct Fine-Tuned on GSM8K with DeepSeek Augmentation

## 🚀 Model Overview

This model is a **fine-tuned version of Qwen2.5-1.5B-Instruct**, optimized for **mathematical problem-solving with step-by-step reasoning**. It was trained on the **GSM8K dataset**, incorporating **Chain-of-Thought (CoT) reasoning** using **DeepSeek augmentation**. 

The model is designed to provide **logical, structured, and interpretable answers**, making it ideal for applications in **education, tutoring, and automated reasoning**.

### 🔹 **Key Features**
- **Base Model:** `Qwen/Qwen2.5-1.5B-Instruct`
- **Fine-Tuned On:** `eagle0504/openai-gsm8k-enhanced-using-together-ai-deepseek-train8k-test1k-v1`
- **Optimized for:** **Mathematical problem-solving & step-by-step logical reasoning**
- **Fine-tuned with:** **LoRA (Low-Rank Adaptation) for efficient memory usage**
- **Inference-ready:** Available on **🤗 Hugging Face** and **compatible with `llama.cpp`**
- **Supports GGUF:** Optimized versions for **Q4_K_M, Q8_0, Q5_K_M, and FP16**

---

## 📂 **Model Details**

- **Developed by:** [Your Name or Organization]
- **Model Type:** Causal Language Model (**Text Generation**)
- **Languages:** English (`en`)
- **License:** MIT License
- **Fine-tuned from:** `Qwen/Qwen2.5-1.5B-Instruct`
- **Training Library:** `transformers` + `unsloth` + `trl`
- **Quantization:** GGUF (`Q4_K_M, Q8_0, Q5_K_M, f16`)

🔗 **Hugging Face Repository**:  
👉 [Fine-tuned Qwen2.5-1.5B-Instruct](https://huggingface.co/your-repo-id)

---

## 🛠 How to Use the Model

### **Using `transformers` in Python**
You can load and use the model with 🤗 `transformers` as follows:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "your-repo-id"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Move model to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# Example inference
question = "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"
inputs = tokenizer(question, return_tensors="pt").to(device)
output = model.generate(**inputs, max_length=200)

# Decode response
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

---

## 🖥️ Running the Model with `llama.cpp` (Mac/Linux/Windows)

The model is **quantized** into GGUF format and can run on Mac **without a GPU** using `llama.cpp`.

### **1️⃣ Install `llama.cpp`**
```sh
brew install llama.cpp
```

### **2️⃣ Download the Model**
```sh
mkdir -p ~/llama_models && cd ~/llama_models
wget https://huggingface.co/your-repo-id/resolve/main/q8_0.gguf
```

### **3️⃣ Run the Model**
```sh
llama-cli -m ~/llama_models/q8_0.gguf --interactive
```

Or you can use the following

```sh
llama-cli -hf eagle0504/qwen-2_5-1_5b-instruct-using-openai-gsm8k-data-enhanced-with-deepseek-v4:Q8_0
```

### **4️⃣ Test with a Prompt**
```sh
llama-cli -m ~/llama_models/q8_0.gguf -p "Explain quantum computing in simple terms."
```

---

## 🏋️ **Training Details**

### **📊 Dataset Used**
The model was fine-tuned on:  
🔹 [`eagle0504/openai-gsm8k-enhanced-using-together-ai-deepseek-train8k-test1k-v1`](https://huggingface.co/datasets/eagle0504/openai-gsm8k-enhanced-using-together-ai-deepseek-train8k-test1k-v1)  

This dataset contains:
- **8K training samples**
- **1K testing samples**
- Features: `"question"`, `"answer"`, `"cot"` (Chain-of-Thought)

### **⚙️ Training Configuration**
- **Framework:** `transformers` + `unsloth` + `trl`
- **Optimization:** 
  - **LoRA (Low-Rank Adaptation)** applied to QKV projections
  - **Learning Rate:** `1e-6`
  - **AdamW Optimizer (8-bit)**
  - **Mixed Precision (`bf16` or `fp16`)**
  - **Batch Size:** `8`
  - **Gradient Accumulation Steps:** `1`
- **Max Sequence Length:** `1024`

---

## 📊 **Model Performance**

### **✅ Training Loss**
| Step | Training Loss | Reward | KL |
|------|--------------|--------|------|
| 1    | 0.0000       | 0.0000 | 0.0000 |
| 500  | 0.0033       | 0.2617 | 0.0821 |
| 1000 | 0.0028       | 0.1359 | 0.0696 |
| 1500 | 0.0062       | 1.3781 | 0.1559 |

### **🧪 Testing & Expected Results**
The model was evaluated on the **1K test samples** and showed strong accuracy in multi-step problem-solving.

Example expected response:
```text
To solve the problem, we first find the clips sold in May:
  Clips in May = 48 / 2 = 24
Next, we find the total:
  Total Clips = 48 + 24 = 72
#### Answer: 72
```

---

## 🚨 **Bias, Risks, and Limitations**
### ⚠️ **Potential Risks**
- May **hallucinate** incorrect reasoning steps if prompts are unclear.
- Could struggle with **complex mathematical problems** outside its training data.
- **Limited generalization** to non-math reasoning tasks.

### 🎯 **Recommendations**
- If using this model for **critical applications**, verify outputs with human review.
- For **better performance**, fine-tune on **larger datasets** with real-world numerical reasoning.

---

## 🌍 **Environmental Impact**
**Estimated Carbon Emissions:**
- **Hardware Used:** NVIDIA A100 GPU
- **Training Time:** ~5 hours
- **Estimated CO2 Emitted:** ~8.2 kg CO2eq (via [ML Impact Calculator](https://mlco2.github.io/impact#compute))

---

## 📖 **Citation**
If you use this model in your research, please cite it as:

```bibtex
@misc{your_model_2024,
  title={Fine-Tuned Qwen2.5-1.5B-Instruct on GSM8K with DeepSeek Augmentation},
  author={Your Name},
  year={2024},
  url={https://huggingface.co/your-repo-id}
}
```

---

## 📩 **Model Card Contact**
For questions, suggestions, or issues, reach out via [Hugging Face Discussions](https://huggingface.co/your-repo-id/discussions).

---

🎉 **Thank you for using this model!** If you find it useful, please ⭐ it on **Hugging Face**! 🚀🔥