docs/MODEL_CARD.md

<p align="center">
  <a href="https://github.com/my-ai-stack/stack-2.9">
    <img src="https://img.shields.io/badge/GitHub-View%20Repo-blue?style=flat-square&logo=github" alt="GitHub">
  </a>
  <a href="https://huggingface.co/spaces/my-ai-stack/stack-2-9-demo">
    <img src="https://img.shields.io/badge/HF%20Space-Demo-green?style=flat-square&logo=huggingface" alt="HuggingFace Space">
  </a>
  <img src="https://img.shields.io/badge/Parameters-1.5B-purple?style=flat-square" alt="Parameters">
  <img src="https://img.shields.io/badge/Context-128K-orange?style=flat-square" alt="Context">
  <img src="https://img.shields.io/badge/License-Apache%202.0-yellow?style=flat-square" alt="License">
</p>

---

# Stack 2.9

> A fine-tuned code assistant built on Qwen2.5-Coder-1.5B, trained on Stack Overflow data

Stack 2.9 is a specialized code generation model fine-tuned from [Qwen/Qwen2.5-Coder-1.5B](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B) on Stack Overflow Q&A data for improved programming assistance.

## Key Features

- **Specialized for Code**: Trained on Stack Overflow patterns for better code generation
- **128K Context**: Handle larger codebases and complex documentation
- **Efficient**: Runs on consumer GPUs (RTX 3060+)
- **Open Source**: Apache 2.0 licensed

---

## Model Details

| Attribute | Value |
|-----------|-------|
| **Base Model** | Qwen/Qwen2.5-Coder-1.5B |
| **Parameters** | 1.5B |
| **Context Length** | 131,072 tokens (128K) |
| **Fine-tuning Method** | LoRA (Rank 8) |
| **Precision** | FP16 |
| **License** | Apache 2.0 |
| **Release Date** | April 2026 |

### Architecture

| Specification | Value |
|--------------|-------|
| Architecture | Qwen2ForCausalLM |
| Hidden Size | 1,536 |
| Num Layers | 28 |
| Attention Heads | 12 (Q) / 2 (KV) |
| GQA | Yes (2 KV heads) |
| Intermediate Size | 8,960 |
| Vocab Size | 151,936 |
| Activation | SiLU (SwiGLU) |
| Normalization | RMSNorm |

---

## Quickstart

### Installation

```bash
pip install transformers>=4.40.0 torch>=2.0.0 accelerate
```

### Code Example

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "my-ai-stack/Stack-2-9-finetuned"

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Chat interface
messages = [
    {"role": "system", "content": "You are Stack 2.9, a helpful coding assistant."},
    {"role": "user", "content": "Write a Python function to calculate fibonacci numbers"}
]

# Apply chat template
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

# Generate
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)

# Decode response
response = tokenizer.decode(
    generated_ids[0][len(model_inputs.input_ids[0]):],
    skip_special_tokens=True
)
print(response)
```

### Interactive Chat

```bash
python chat.py
```

---

## Training Details

| Specification | Value |
|--------------|-------|
| **Method** | LoRA (Low-Rank Adaptation) |
| **LoRA Rank** | 8 |
| **LoRA Alpha** | 16 |
| **Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **Epochs** | ~0.8 |
| **Final Loss** | 0.0205 |
| **Data Source** | Stack Overflow Q&A |

### Training Data

Fine-tuned on Stack Overflow code Q&A pairs including:
- Python code solutions and snippets
- Code explanations and documentation
- Programming patterns and best practices
- Bug fixes and debugging examples
- Algorithm implementations

---

## Evaluation

| Benchmark | Score | Notes |
|-----------|-------|-------|
| **HumanEval** | ~35-40% | Based on base model benchmarks |
| **MBPP** | ~40-45% | Python-focused evaluation |

> **Note**: Full benchmark evaluation is in progress. The model inherits strong coding capabilities from Qwen2.5-Coder and is specialized for Stack Overflow patterns.

---

## Hardware Requirements

| Configuration | GPU | VRAM |
|---------------|-----|------|
| FP16 | RTX 3060+ | ~4GB |
| 8-bit | RTX 3060+ | ~2GB |
| 4-bit | Any modern GPU | ~1GB |
| CPU | None | ~8GB RAM |

---

## Capabilities

- **Code Generation**: Python, JavaScript, TypeScript, SQL, Go, Rust, and more
- **Code Completion**: Functions, classes, and entire snippets
- **Debugging**: Identify and fix bugs with explanations
- **Code Explanation**: Document and explain code behavior
- **Programming Q&A**: Answer technical questions

---

## Limitations

- **Model Size**: At 1.5B parameters, smaller than state-of-the-art models (7B+)
- **Training Data**: Python-heavy; other languages may have lower quality
- **Hallucinations**: May occasionally generate incorrect code; verification recommended
- **Tool Use**: Base model without native tool-calling (see enhanced version)

---

## Comparison

| Feature | Qwen2.5-Coder-1.5B | Stack 2.9 |
|---------|-------------------|-----------|
| Code Generation | General | Stack Overflow patterns |
| Python Proficiency | Baseline | Enhanced |
| Context Length | 128K | 128K |
| Specialization | General code | Stack Overflow Q&A |

---

## Citation

```bibtex
@misc{my-ai-stack/stack-2-9-finetuned,
  author = {Walid Sobhi},
  title = {Stack 2.9: Fine-tuned Qwen2.5-Coder-1.5B on Stack Overflow Data},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/my-ai-stack/Stack-2-9-finetuned}
}
```

---

## Related Links

- [GitHub Repository](https://github.com/my-ai-stack/stack-2.9)
- [HuggingFace Space Demo](https://huggingface.co/spaces/my-ai-stack/stack-2-9-demo)
- [Base Model](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B)
- [Qwen2.5-Coder-7B](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct)
- [Qwen2.5-Coder-32B](https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct)

---

## License

Licensed under the Apache 2.0 license. See [LICENSE](LICENSE) for details.

---

*Model Card Version: 2.0*
*Last Updated: April 2026*