SoarAILabs/KiteResolve-20B

🪁 KiteResolve-20B: AI-Powered Merge Conflict Resolution

Developed by [Soar AI Labs](https://huggingface.co/SoarAILabs)

<div align="center"> <img src="https://img.shields.io/badge/License-Apache%202.0-blue.svg" alt="License"> <img src="https://img.shields.io/badge/Model-20B%20Parameters-red.svg" alt="Parameters"> <img src="https://img.shields.io/badge/Task-Code%20Generation-green.svg" alt="Task"> <img src="https://img.shields.io/badge/BLEU-54.83-orange.svg" alt="BLEU Score"> </div>

🚀 Model Description

KiteResolve-20B is a fine-tuned version of GPT-OSS-20B specifically engineered for automated Git merge conflict resolution. This model transforms the tedious process of manually resolving merge conflicts into an intelligent, automated workflow that understands code semantics across multiple programming languages.

✨ Key Features

• —🎯 20% Exact Match Accuracy on real-world merge conflicts
• —📈 12% Token-F1 Score Improvement over base model
• —🌐 Multi-Language Support: Java, JavaScript, Python, C#, TypeScript, and more
• —⚡ Fast Inference: Optimized for CLI and webhook integrations
• —🔧 Production Ready: Designed for enterprise Git workflows

📊 Performance Metrics

Evaluated on 50 held-out samples from real-world merge conflicts.

🛠️ Usage

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
from unsloth.chat_templates import get_chat_template

# Load the model
model = AutoModelForCausalLM.from_pretrained("SoarAILabs/KiteResolve-20B")
tokenizer = AutoTokenizer.from_pretrained("SoarAILabs/KiteResolve-20B")
tokenizer = get_chat_template(tokenizer, chat_template="gpt-oss")

# Resolve a merge conflict
conflict = """
<<<<<<< ours
function calculateTotal(items) {
    return items.reduce((sum, item) => sum + item.price, 0);
}
=======
function calculateTotal(items) {
    return items.map(item => item.price).reduce((a, b) => a + b, 0);
}
>>>>>>> theirs
"""

messages = [{"role": "user", "content": f"Resolve this merge conflict:\n```{conflict}```"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer([prompt], return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200, do_sample=False)

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

Ollama 🦙️

ollama run hf.co/SoarAILabs/KiteResolve-20B/model-q4_k_m.gguf