Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch

Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch

A LoRA fine-tuned model based on Kwaipilot/KAT-Dev specialized for the Hyperswitch Rust codebase. This model excels at understanding payment processing patterns, Hyperswitch architecture, and Rust development practices.

🎯 Model Description

This LoRA adapter was trained on 16,731 samples extracted from the Hyperswitch codebase to enhance code understanding, explanation, and generation within the payment processing domain.

• Base Model: Kwaipilot/KAT-Dev
• Training Type: Causal Language Modeling (CLM) with LoRA
• Domain: Payment Processing, Rust Development
• Specialization: Hyperswitch codebase patterns and architecture

🏆 RustEvo Benchmark

<p float="center"> <img src="https://huggingface.co/AdityaNarayan/Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch/resolve/main/bench_asset/rustEvo_bench_chart.png" width="600" /> <img src="https://huggingface.co/AdityaNarayan/Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch/resolve/main/bench_asset/rustEvo_bench_table.png" width="600" /> </p>

RustEvo Github Repo : https://github.com/SYSUSELab/RustEvo

📊 Training Details

Dataset Composition

• Total Samples: 16,731
  • File-level samples: 2,120 complete files
  • Granular samples: 14,611 extracted components
    • Functions: 4,121
    • Structs: 5,710
    • Traits: 223
    • Implementations: 4,296
    • Modules: 261

LoRA Configuration

r: 64                   # LoRA rank
alpha: 128              # LoRA alpha (2*r)
dropout: 0.05           # LoRA dropout
target_modules:         # Applied to all linear layers
  - q_proj, k_proj, v_proj, o_proj
  - gate_proj, up_proj, down_proj

Training Hyperparameters

• Epochs: 5
• Batch Size: 2 per device (16 effective with gradient accumulation)
• Learning Rate: 5e-5 (cosine schedule)
• Max Context: 8,192 tokens
• Hardware: 2x NVIDIA H200 (80GB each)
• Training Time: ~4 hours (2,355 steps)

Training Results

Final Loss: 0.48 (from 1.63)
Perplexity: 1.59 (from 5.12)  
Accuracy: 89% (from 61%)

🚀 Usage

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Kwaipilot/KAT-Dev",
    dtype=torch.bfloat16,
    device_map="auto"
)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Kwaipilot/KAT-Dev")
# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "AdityaNarayan/Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch")
# Generate code
prompt = """// Hyperswitch payment processing
pub fn validate_payment_method("""
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.2,  # Lower temperature for code generation
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended Settings

• Temperature: 0.2-0.3 for code generation
• Temperature: 0.5-0.7 for explanations and documentation
• Max tokens: 1024 for most tasks

🛠️ Technical Specifications

• Context Window: 8,192 tokens
• Precision: bfloat16
• Memory Usage: ~78GB VRAM (32B base model)
• Inference Speed: Optimized with Flash Attention 2

🙏 Acknowledgments

• Kwaipilot Team for the excellent KAT-Dev base model
• Hyperswitch Team for the open-source payment processing platform
• Hugging Face for the transformers and PEFT libraries

📞 Citation

@misc{hyperswitch-kat-dev-lora-2024,
  title={Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch},
  author={Aditya Narayan},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/AdityaNarayan/Kwaipilot-KAT-Dev-CPT-LoRA-Adapter-HyperSwitch}
}