Anni

   

Anni is a high-performance code assistant built on the Qwen3 14B architecture.
Fine-tuned on OpenCodeReasoning-2, it is engineered to excel in deep algorithmic reasoning, complex data structure implementation, and competitive programming.

View Demo • Quick Start • Benchmarks • Training

---

✨ Key Features

Feature	Description
🧠 Deep Reasoning	Optimized for hard logic puzzles and algorithmic challenges.
⚡ High Efficiency	Supports vLLM serving and GGUF for consumer hardware.
📚 Large Context	32k context window for processing extensive codebases.
🛠️ Dev Ready	Comes with full training scripts, merging tools, and a web UI.

---

🎥 Demo

demo.mp4

Anni solving a hard-difficulty LeetCode problem in real-time (1x speed on a single L40 GPU)

---

🚀 Quick Start

Experience Anni immediately without local setup using Google Colab.

method	link	description
GGUF (Recommended)		Run standard inference on free tier GPUs.
vLLM Serving		High-throughput serving using vLLM.

---

📊 Benchmarks

Anni was evaluated on LiveCodeBench (LCB), demonstrating superior performance in code generation and reasoning tasks compared to base models.

---

🛠️ Development Setup

If you wish to fine-tune or run Anni locally, follow these steps.

1. Prerequisites

Ensure tmux is installed.

pip install -r requirements.txt

2. Configuration

Set up your environment variables for WandB, HuggingFace, and ModelScope.

mv config/example.env config/.env
# Open config/.env and paste your API keys

Edit config/config.yaml to adjust hyperparameters.

Note: Specify the LOCAL_STORAGE_PATH in src/train.py before starting.

3. Training

Launch the training pipeline:

./scripts/train.sh

📂 Project Structure

Anni/
├── config/                 # Configuration files
│
├── scripts/                # Shell scripts for automation
│   ├── train.sh            # Start training pipeline
│   ├── eval.sh             # Run LiveCodeBench evaluation
│   ├── serve.sh            # Spin up vLLM server
│   └── terminate_train.sh  # Kill training processes
│
├── src/                    # Python source code
│   ├── preprocess.py       # Downloads & preps OpenCodeReasoning-2
│   ├── train.py            # Main fine-tuning logic
│   ├── save.py             # Merges LoRA adapters (BF16 & GGUF)
│   ├── inference.py        # Run inference with the fine-tuned model
│   ├── upload.py           # Pushes to HF/ModelScope
│   └── utils/              # Utility functions
│
└── web/                    # Frontend Interface

👉 View Frontend Documentation

⚖️ License & Disclaimer

License

• Model Weights & Training Code: Released under the MIT License.

• Trademarks: The project name (Anni), assets, and frontend code are trademarks of the owner (Hans) and may not be used without explicit permission.

Dataset Attribution

• Trained on OpenCoderReasoning-2 (CC-BY-4.0).

Disclaimer: This model may generate incorrect or unsafe code. Evaluate and verify outputs before using in production environments.