Chitti Beast 🔥

Elite coding assistant fine-tuned for professional developers and security researchers.

Overview

Chitti is a 7B parameter coding assistant fine-tuned from Qwen2.5-Coder-7B-Instruct with a curated dataset combination and abliteration applied for unrestricted security research use.

Built entirely on MoLab (Marimo) — RTX PRO 6000 Blackwell, 102GB VRAM.

Benchmark Results

Benchmark	Chitti	Reference	Status
HumanEval+	76.83%	DeepSeek-Coder-7B: 73%	✅ Beats reference
MultiPL-E avg	67.49%	Qwen2.5-Coder-7B: 65%	✅ Beats reference
LiveCodeBench	23.5%	~28%	✅ Credible

Training Pipeline

• Base: Qwen2.5-Coder-7B-Instruct
• Abliteration: Refusal direction removal across all 28 layers (alpha=2.0)
• Fine-tuning: LoRA on OpenCodeInstruct via Unsloth + TRL
• Merge: LoRA fused into abliterated base — single clean model
• Parameters: 7.62B

Capabilities

• Strong algorithmic coding (beats DeepSeek-Coder-7B on HumanEval+)
• Multilingual coding — Python, Java, C++, Rust, C#, Go (beats base model)
• Security-focused response format: Warning → Code → Detection/Cleanup
• No refusals on legitimate security research tasks

Tech Stack

Python PyTorch Unsloth HuggingFace PEFT TRL Gradio

Access

This model is available for enterprise licensing and partnerships.

📧 kkishann4@gmail.com
💼 LinkedIn
🐙 GitHub

---

Built by Kishan N — B.Tech CSE, India