KAT-Dev-CPT-LoRA-HS-32K-maxToken-v1

A specialized LoRA fine-tuned adapter built on top of Kwaipilot/KAT-Dev (32B), designed for deep understanding of the Hyperswitch (Rust) payment orchestration codebase. This model uses a 3-phase curriculum training pipeline, progressively enhancing the model's grasp of Rust patterns, PR changes, repository structure, and payment processing logic.


πŸš€ Overview

This LoRA adapter was trained with a phased CPT strategy:

Phase 1 β€” Foundation

Learns core repository structure, Rust syntax, basic modules, and Hyperswitch architectural patterns.

Phase 2 β€” Evolution

Exposes the model to progressively complex components, multi-file interactions, workflows, and feature evolution.

Phase 3 β€” PR Mastery

Specializes on real PR changes, diffs, refactors, and reasoning across multi-module changes.

The final result is a high-signal Rust-aware, Hyperswitch-specialized LoRA adapter ideal for:

  • Code generation
  • Code explanation
  • PR reasoning
  • Diff summarization
  • Documentation generation
  • Rust workflow automation

πŸ”§ Training Details

LoRA Configuration

r: 128
alpha: 256
dropout: 0.05
target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

Hyperparameters

learning_rate: 1e-4
micro_batch_size: 1
gradient_accumulation_steps: 6
sequence_length: 32768
train/val split: 95/5
precision: bf16

Hardware

num_gpus: 8
gpu_name: NVIDIA H200

πŸ“Š Phased Training Metrics

Phase 1 β€” Foundation

Dataset: phase1_foundation.jsonl
Epochs: 3

Metric Train Eval
Loss 0.2918 0.2434
Entropy 0.2052 0.2355
Mean Token Accuracy 0.9505 0.9331
Perplexity β€” 1.2756
Tokens 8.88M 8.88M

Phase 2 β€” Evolution

Dataset: phase2_evolution.jsonl
Epochs: 2

Metric Train Eval
Loss 0.7255 0.7661
Entropy 0.5080 0.7210
Mean Token Accuracy 0.8641 0.8110
Perplexity β€” 2.1514
Tokens 23.48M 23.48M

Phase 3 β€” PR Mastery

Dataset: phase3_pr_mastery.jsonl
Epochs: 2

Metric Train Eval
Loss 0.5378 0.5606
Entropy 0.4781 0.5254
Mean Token Accuracy 0.8749 0.8569
Perplexity β€” 1.7516
Tokens 15.45M 15.45M

πŸ“ˆ Summary Across All Phases

total_epochs: 7
total_phases: 3

initial_train_loss: 0.2918
final_train_loss: 0.5378

initial_eval_loss: 0.2434
final_eval_loss: 0.5606

initial_perplexity: 1.2756
final_perplexity: 1.7516

πŸ™ Acknowledgments

  • Kwaipilot Team β€” For the excellent KAT-Dev 32B base model
  • Juspay / Hyperswitch β€” For the rich open-source Rust codebase
  • Hugging Face β€” For PEFT, TRL, and Transformers

πŸ“š Citation

@misc{katdev-hyperswitch-phasedlora-2025,
  title={KAT-Dev-CPT-LoRA-HS-32K-maxToken-v1},
  author={Aditya Narayan},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/AdityaNarayan/KAT-Dev-CPT-LoRA-HS-32K-maxToken-v1}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for AdityaNarayan/KAT-Dev-CPT-LoRA-HS-32K-maxToken-v1

Base model

Kwaipilot/KAT-Dev
Finetuned
(4)
this model

Dataset used to train AdityaNarayan/KAT-Dev-CPT-LoRA-HS-32K-maxToken-v1