SP-TransCoder: Python to C++ Code Translation

Course: CS 175: Project in Artificial Intelligence, Winter 2026

Author/Team: Sterling Radisay, Elian Hijmans Malessy, Sia Aggarwal

This repository contains the software, models, and evaluation code for SP-TransCoder, a project exploring artificial intelligence techniques for translating code between Python and C++.

Libraries Used

• torch (PyTorch) - Deep learning framework for model training and inference.
• transformers (Hugging Face) - Model architecture (T5ForConditionalGeneration) and tokenization.
• datasets (Hugging Face) - Downloading and streaming training/evaluation data.
• codebleu - Calculating syntactic and semantic similarity metrics.
• tqdm - Progress bars for training and evaluation loops.

Publicly Available Codes Used

• Salesforce/codet5-small (https://huggingface.co/Salesforce/codet5-small): Used as the base pre-trained LLM architecture and tokenizer. Unmodified.
• bigcode/the-stack-smol (https://huggingface.co/datasets/bigcode/the-stack-smol): Unpaired monolingual code dataset used for Unsupervised Back-Translation. Unmodified.
• iNeil77/CodeNet (https://huggingface.co/datasets/iNeil77/CodeNet): Dataset used to mine paired Python/C++ code. Unmodified.
• THUDM/humaneval-x (https://huggingface.co/datasets/THUDM/humaneval-x): Dataset used for execution-based unit testing evaluation. Unmodified.

Scripts/Functions Written by Our Team

All original source code for this project is located within the src/ directory.

Data Preparation & Loading

• src/dataset.py: Central data loader containing classes for unpaired (The Stack), paired (CodeNet), and snippet (XLCoST) datasets (approx. 110 lines).
• src/prepare_paired_data.py: Utility script that mines the CodeNet dataset to align "Accepted" Python and C++ solutions matching the same problem ID (approx. 70 lines).

Model Trainers

• src/models/supervised_trainer.py: Algorithm for standard Supervised Fine-Tuning on paired datasets (approx. 80 lines).
• src/models/backtranslation_trainer.py: Algorithm for Unsupervised Iterative Back-Translation using noisy generation (approx. 90 lines).
• src/models/denoising_trainer.py: Algorithm for Denoising Auto-Encoding (DAE) via token dropping, masking, and shuffling (approx. 110 lines).

Training Pipelines

• src/train_supervised.py: Main execution loop for supervised training (approx. 40 lines).
• src/train_backtranslation.py: Main execution loop for back-translation (approx. 30 lines).
• src/train_semi_supervised.py: Main execution loop alternating supervised and back-translation phases (approx. 50 lines).
• src/train_semi_supervised_with_denoising.py: Main execution loop running a 3-phase cycle of DAE, supervised, and back-translation (approx. 60 lines).

Benchmarking & Evaluation (src/bench/)

• src/bench/evaluate.py: Calculates the CodeBLEU score for models using tokenized parallel data (approx. 160 lines).
• src/bench/naive_copy.py: Establishes a "do-nothing" baseline CodeBLEU score (approx. 60 lines).
• src/bench/unit_testing.py: Generates execution-based translations for the CodeGeeX evaluation harness (approx. 80 lines).
• src/bench/visual_inspection.py: Performs qualitative sanity checks on model checkpoints using basic code snippets (approx. 70 lines).
• src/eval.sh: Bash script to automate batched evaluation across multiple saved checkpoints (approx. 40 lines).