| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: Qwen/Qwen2.5-Coder-7B-Instruct |
| pipeline_tag: text-generation |
| library_name: transformers |
| tags: |
| - code |
| - coding-assistant |
| - security |
| - qwen2.5 |
| - fine-tuned |
| --- |
| |
| # 🚀 Chitti |
|
|
| > **Enterprise coding assistant for secure software development and cybersecurity research.** |
|
|
| --- |
|
|
| # Overview |
|
|
| Chitti is a fine-tuned 7B coding assistant designed for professional developers, security researchers, and software engineering teams. |
|
|
| Built upon **Qwen2.5-Coder-7B-Instruct**, Chitti combines abliteration and LoRA fine-tuning to improve coding performance while remaining practical for legitimate cybersecurity and software engineering workflows. |
|
|
| The complete training, evaluation, and deployment pipeline was developed independently using **MoLab (Marimo)**. |
|
|
| --- |
|
|
| # Highlights |
|
|
| - 🚀 Fine-tuned from Qwen2.5-Coder-7B-Instruct |
| - 🛡 Built for secure software development |
| - 🧠 Optimized for professional coding workflows |
| - 🌍 Multilingual programming support |
| - 📈 Benchmarked against leading open-source coding models |
|
|
| --- |
|
|
| # Benchmark Results |
|
|
| | Benchmark | Chitti | Reference | Result | |
| |-----------|--------|-----------|--------| |
| | HumanEval+ | **76.83%** | DeepSeek-Coder-7B (73%) | ✅ | |
| | MultiPL-E (Average) | **67.49%** | Qwen2.5-Coder-7B (65%) | ✅ | |
| | LiveCodeBench | **23.5%** | Competitive Baseline | ✅ | |
|
|
| --- |
|
|
| # Training Pipeline |
|
|
| ```text |
| Qwen2.5-Coder-7B-Instruct |
| │ |
| ▼ |
| Abliteration |
| │ |
| ▼ |
| LoRA Fine-tuning |
| │ |
| ▼ |
| Model Merge |
| │ |
| ▼ |
| Benchmark Evaluation |
| │ |
| ▼ |
| Chitti |
| ``` |
|
|
| ### Base Model |
|
|
| - Qwen2.5-Coder-7B-Instruct |
|
|
| ### Fine-tuning |
|
|
| - LoRA |
| - Unsloth |
| - TRL |
| - PEFT |
|
|
| ### Parameters |
|
|
| **7.62 Billion** |
|
|
| --- |
|
|
| # Capabilities |
|
|
| - Algorithmic problem solving |
| - Code generation |
| - Code explanation |
| - Code refactoring |
| - Debugging assistance |
| - Software engineering support |
| - Security-oriented programming workflows |
| - Multilingual coding (Python, Java, C++, Go, Rust, C# and more) |
|
|
| --- |
|
|
| # Intended Use |
|
|
| Chitti is designed for: |
|
|
| - Professional software development |
| - Cybersecurity research |
| - Secure coding assistance |
| - Learning and education |
| - Code review |
| - Algorithm practice |
| - Development productivity |
|
|
| --- |
|
|
| # Tech Stack |
|
|
| - Python |
| - PyTorch |
| - Hugging Face Transformers |
| - Unsloth |
| - PEFT |
| - TRL |
| - LoRA |
|
|
| --- |
|
|
| # Hardware |
|
|
| Training and evaluation were performed on **MoLab (Marimo)** using NVIDIA RTX PRO 6000 Blackwell GPUs. |
|
|
| --- |
|
|
| ## Availability |
|
|
| Chitti is not currently available for public distribution. |
|
|
| Enterprise deployments, licensing, and collaboration opportunities are available through direct engagement. |
|
|
| --- |
|
|
| # Contact |
|
|
| 📧 **Email:** kkishann4@gmail.com |
|
|
| 💼 **LinkedIn:** https://linkedin.com/in/kkishann |
|
|
| 🐙 **GitHub:** https://github.com/Ki1shan |
|
|
|
|
| --- |
|
|
| # License |
|
|
| Apache License 2.0 |
|
|
| --- |
|
|
| # Author |
|
|
| **Kishan N** |
|
|
| Cybersecurity Engineer + AI Systems Builder |
|
|
| --- |
|
|
| > **Building AI systems for secure software engineering and cybersecurity research.** |
