README.md

---
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
base_model_relation: finetune
library_name: transformers
language:
- en
tags:
- code
- codeqwen
- chat
- qwen
- qwen-coder
license: gpl-3.0
datasets:
- bunyaminergen/Stable-Code-Python-SFT
pipeline_tag: text-generation
license_link: https://huggingface.co/bunyaminergen/Qwen2.5-Coder-1.5B-Instruct-SFT/blob/main/LICENSE
---

# Qwen2.5-Coder-1.5B-Instruct-SFT

The Qwen2.5-Coder-1.5B-Instruct-SFT model has been fine-tuned using Supervised Fine Tuning (SFT) method specifically to
enhance its performance in generating docstrings, a critical task in coding. Currently, it generates docstrings
following the Google style guide. For differences compared to the base (pre-trained) model, please refer to
the [Comparison](#comparison) section.

**Note**: _Since the model has only been supervised fine-tuned (SFT) for roughly 11 epochs and hasn't been enhanced with
reinforcement learning from human feedback (RLHF), it mostly generates docstrings in the Google Docstring style.
However, it might occasionally produce them in a different style. In such cases, you may need to regenerate the
docstrings once or multiple times._

**Note**: _If you would like to contribute to this repository,
please read the [CONTRIBUTING](CONTRIBUTING.md) first._

---

### TableofContents

- [Usage](#usage)
- [Comparison](#comparison)
- [Dataset](#dataset)
- [Training](#training)
- [Documentations](#documentations)
- [License](#licence)
- [Links](#links)
- [Team](#team)
- [Contact](#contact)
- [Reference](#reference)
- [Citation](#citation)

---

### Usage

#### Hugging Face

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

model_name = "bunyaminergen/Qwen2.5-Coder-1.5B-Instruct-SFT"

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4"
)

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=quant_config,
    device_map="auto"
)

model.eval()

messages = [
    {"role": "system", "content": "You are a senior Python developer."},
    {"role": "user", "content": "Give me a quick example of bubble sort in Python."}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

max_new_tokens = 512
temperature = 0.9

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        do_sample=True,
        top_p=0.95,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id
    )

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

print(result)
```

---

### Comparison

#### Base Model

`System:` You are a senior Python developer..

`User:` Give me a quick an example of bubble sort in Python.

`Assistant:` Bubble Sort is a simple sorting algorithm that repeatedly steps through the list, compares adjacent
elements and swaps them if they are in the wrong order. The pass through the list is repeated until the list is sorted.

Here's a basic implementation of Bubble Sort in Python:

```python
def bubble_sort(arr):
    n = len(arr)
    # Traverse through all array elements
    for i in range(n):
        # Last i elements are already in place
        for j in range(0, n - i - 1):
            # Swap if the element found is greater than the next element
            if arr[j] > arr[j + 1]:
                arr[j], arr[j + 1] = arr[j + 1], arr[j]


# Example usage:
arr = [64, 34, 25, 12, 22, 11, 90]
bubble_sort(arr)
print("Sorted array:", arr)
```

In this example:

- We define a function `bubble_sort` that takes an array as input.
- Inside the function, we use two nested loops to iterate over each pair of adjacent elements in the array.
- If the current element (`arr[j]`) is greater than the next element (`arr[j+1]`), we swap them.
- This process continues until no more swaps are needed, indicating that the array is sorted.
- Finally, we demonstrate how to use the `bubble_sort` function by sorting an array and printing the sorted result.

Bubble Sort has a time complexity of O(n^2) in the worst case scenario, making it less efficient than other sorting
algorithms like QuickSort or MergeSort. However, it is easy to understand and implement, which makes it suitable for
small data sets or educational purposes.

#### Supervised Fine Tuned Model

`System:` You are a senior Python developer.

`User:` Give me a quick an example of bubble sort in Python.

`Assistant:`  Here's a simple implementation of the Bubble Sort algorithm in Python, along with its docstring following
Google Python Style Guide:

```python
def bubble_sort(arr):
    """
    Bubble sorts an array of integers in ascending order.

    Args:
        arr (list): A list of integers to be sorted.

    Returns:
        list: The sorted list of integers.

    Raises:
        TypeError: If the input is not a list or contains non-integer elements.

    Examples:
        >>> bubble_sort([5, 3, 1, 4, 6])
        [1, 3, 4, 5, 6]
    """

    # Check if the input is a list and contains only integers
    if not isinstance(arr, list) or not all(isinstance(x, int) for x in arr):
        raise TypeError("Input must be a list of integers.")

    n = len(arr)

    # Traverse through all array elements
    for i in range(n):

        # Last i elements are already in place
        for j in range(0, n - i - 1):

            # Traverse the array from 0 to n-i-1
            # Swap if the element found is greater than the next element
            if arr[j] > arr[j + 1]:
                arr[j], arr[j + 1] = arr[j + 1], arr[j]

    return arr
```

This implementation follows the Bubble Sort algorithm, which repeatedly steps through the list, compares adjacent
elements, and swaps them if they are in the wrong order. The pass through the list is repeated until the list is sorted.
This implementation includes error handling and examples for usage.

---

### Dataset

- [bunyaminergen/Stable-Code-Python-SFT](https://huggingface.co/datasets/bunyaminergen/Stable-Code-Python-SFT)

---

### Training

#### Hyperparameters

| Hyperparameter              | Value                                 |
|-----------------------------|---------------------------------------|
| Base Model                  | `Qwen/Qwen2.5-Coder-1.5B-Instruct`    |
| Fine-tuning Method          | QLoRA (Quantized Low-Rank Adaptation) |
| Task Type                   | `CAUSAL_LM`                           |
| Number of Epochs            | `11`                                  |
| Batch Size                  | `8`                                   |
| Gradient Accumulation Steps | `2`                                   |
| Effective Batch Size        | `16` (8 × 2)                          |
| Learning Rate               | `1e-4`                                |
| Optimizer                   | `AdamW`                               |
| Precision                   | `BF16 Mixed Precision`                |
| Evaluation Strategy         | `None`                                |
| Max Sequence Length         | `1024 tokens`                         |
| Logging Steps               | every `1000` steps                    |
| Save Checkpoint Steps       | every `7200` steps                    |
| Output Directory            | Overwritten per run                   |
| Experiment Tracking         | `MLflow` (local tracking)             |
| Experiment Name             | `AssistantFineTuning`                 |
| MLflow Run Name             | `AssistantFT`                         |

#### PEFT (QLoRA) Configuration

| Parameter       | Value                    |
|-----------------|--------------------------|
| LoRA Rank (`r`) | `16`                     |
| LoRA Alpha      | `32`                     |
| LoRA Dropout    | `0.05`                   |
| Target Modules  | `all-linear`             |
| Modules Saved   | `lm_head`, `embed_token` |

#### Dataset

- **Train/Test Split:** `90%/10%`
- **Random Seed:** `19`
- **Train Batched:** `True`
- **Eval Batched:** `True`

#### Tokenizer Configuration

- **Truncation:** Enabled (`max_length=1024`)
- **Masked Language Modeling (MLM):** `False`

#### Speeds, Sizes, Times

- **Total Training Time:** ~11 hours
- **Checkpoint Frequency:** every `7200` steps
- **Checkpoint Steps:**
    - `checkpoint-7200`
    - `checkpoint-14400`
    - `checkpoint-21600`
    - `checkpoint-28800`
    - `checkpoint-36000`
    - `checkpoint-39600` *(final checkpoint)*

#### Compute Infrastructure

**Hardware:**

- GPU: **1 × NVIDIA L40S (48 GB VRAM)**
- RAM: **62 GB**
- CPU: **16 vCPU**

**Software:**

- OS: **Ubuntu 22.04**
- Frameworks: **PyTorch 2.4.0**
- CUDA Version: **12.4.1**

---

### Documentations

- [CONTRIBUTING](CONTRIBUTING.md)

---

### Licence

- [LICENSE](LICENSE)

---

### Links

- [Github](https://github.com/bunyaminergen/)
- [Website](https://bunyaminergen.com)
- [Linkedin](https://www.linkedin.com/in/bunyaminergen)

---

### Team

- [Bunyamin Ergen](https://www.linkedin.com/in/bunyaminergen)

---

### Contact

- [Mail](mailto:[email protected])

---

### Reference

- This model has been fine-tuned using Supervised Fine Tuning (SFT) method from the original
  model [Qwen/Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct).

---

### Citation

```bibtex
@software{       Qwen2.5-Coder-1.5B-Instruct-SFT,
  author       = {Bunyamin Ergen},
  title        = {{Qwen2.5-Coder-1.5B-Instruct-SFT}},
  year         = {2025},
  month        = {04},
}
```

---