README.md

---
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**, designed for **mathematical problem-solving and structured reasoning**. It is trained on an **enhanced GSM8K dataset** incorporating **Chain-of-Thought (CoT) reasoning** augmented by **DeepSeek AI**.

### Key Features
- **Base Model:** Qwen2.5-1.5B-Instruct
- **Fine-Tuned On:** GSM8K enhanced with DeepSeek-V3
- **Optimized for:** Logical problem-solving and math reasoning
- **Fine-tuning method:** LoRA (Low-Rank Adaptation)
- **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:** [Yiqiao Yin](https://www.y-yin.io/)
- **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/eagle0504/qwen-2_5-1_5b-instruct-using-openai-gsm8k-data-enhanced-with-deepseek-v2)

## How to Use the Model

### Using `transformers` in Python
You may need to install `bitsandbytes` by using

```bash
! pip install -U bitsandbytes
```

Then you can use the following code to run inference.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "eagle0504/qwen-2_5-1_5b-instruct-using-openai-gsm8k-data-enhanced-with-deepseek-v2"
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`

### Step 1: Install `llama.cpp`
```sh
brew install llama.cpp
```

### Step 2: Download the Model
```sh
mkdir -p ~/llama_models && cd ~/llama_models
wget https://huggingface.co/eagle0504/qwen-2_5-1_5b-instruct-using-openai-gsm8k-data-enhanced-with-deepseek-v2/resolve/main/q8_0.gguf
```

### Step 3: 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-v2:Q8_0
```

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

## Training Details

### Custom Reward

```python
def count_xml(text: str) -> float:
    """
    Calculates a reward based on the occurrence of certain XML tags and subtracts penalties for content after closing tags.

    Args:
    text (str): The text string to analyze for XML tag consistency.

    Returns:
    float: Total reward score based on XML tag occurrence and penalties.
    """
    count = 0.0
    if text.count("<think>\n") == 1:
        count += 0.125
    if text.count("\n</think>\n") == 1:
        count += 0.125
    if text.count("\n<answer>\n") == 1:
        count += 0.125
        count -= len(text.split("\n</answer>\n")[-1])*0.001
    if text.count("\n</answer>") == 1:
        count += 0.125
        count -= (len(text.split("\n</answer>")[-1]) - 1)*0.001

    # Ensure `<think>` and `</think>` exist
    if "<think>" in text and "</think>" in text:
        count += 1.0  # Higher weight to ensure reasoning consistency
    else:
        count -= 1.0  # Penalize if missing

    return count
```

Each component contributes to the total reward **if conditions are met**:

| Condition | Reward |
|-----------|--------|
| `"<think>\n"` appears exactly **once** | **+0.125** |
| `"\n</think>\n"` appears exactly **once** | **+0.125** |
| `"\n<answer>\n"` appears exactly **once** | **+0.125** |
| `"\n</answer>"` appears exactly **once** | **+0.125** |
| Both `<think>` and `</think>` exist anywhere | **+1.0** |
| No extra text after `"</answer>"` | **No penalty** |

Total possible reward **before penalties**:
\[
0.125 + 0.125 + 0.125 + 0.125 + 1.0 = 1.5
\]

**Potential Penalties**
The function applies penalties for **extra content after `"</answer>"`**:
\[
-\left( \text{length of extra text} \times 0.001 \right)
\]
If the **best case** occurs (i.e., **no extra content**), then:
- **Penalty = 0**  
- **Final Reward = 1.5 (no deductions)**

---

**Best Possible Input Example**
This **ideal input** gives the highest possible reward:

```xml
<think>
Valid reasoning goes here.
</think>

<answer>
Correct final answer here.
</answer>
```

This means we customize the reward function so that we encourage the answer to have reasoning inside. We also know mathematically what the reward should be so we can monitor it during training process.

### 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 applied to QKV projections
- **Learning Rate:** `1e-6`
- **AdamW Optimizer (8-bit)**
- **Mixed Precision (`bf16` or `fp16`)**
- **Batch Size:** `8`
- **Max Sequence Length:** `1024`

## Model Performance

### Training Loss
| Step | XML Count |
|------|-----------|
| 10   | -1 |
| 100  | -1 |
| 500  | -0.6421 |
| 750  | 0.7611 |
| 1000 | 1.0506 |

As we can see, after 1000 steps, we see the reward of XML Count is above 1, which is getting good. This took about `1h 46min 50s` on a T4 GPU in Colab with High RAM. 

## 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

Upcoming

## Contact
For questions, suggestions, or issues, reach out via [Hugging Face Discussions](https://huggingface.co/eagle0504/qwen-2_5-1_5b-instruct-using-openai-gsm8k-data-enhanced-with-deepseek-v1/discussions).

---

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