UniGenCoder: Merging SEQ2SEQ and SEQ2TREE Paradigms for Unified Code Generation

This is the official PyTorch implementation for the following ICSE2025 NIER paper:

Title: UniGenCoder: Merging SEQ2SEQ and SEQ2TREE Paradigms for Unified Code Generation

Our implementation is built on the source code from CodeT5 and Tranx. Thanks for their work.

Environment

We recommend readers to refer to envs.yaml(conda export) or envs.txt(pip export) for more detailed environment information.

Dataset

--task	--sub_task	Description
code generation	nl-java	text-to-code generation on Concode data
code translation	cs-java	code-to-code translation from C# to Java

Download

• data and checkpoints (under preparing)

Quick Start

1. Before starting, please place data in the right position. The correct data directory structure should be:

   data
   ├── concode
       ├── train.json
       ├── dev.json
       └── test.json

2. Generate the grammar file.

   cd UniGenCoder
   python build_ast.py

3. Prepare the seq2seq teacher and seq2tree teacher and get corresponding best model using checkpoint average:

   cd CodeT5
   bash sh/fine_tune_concode.sh
   bash sh/average_model.sh
   bash sh/fine_tune_concode_tree.sh
   bash sh/average_model.sh

4. UniGenCoder backbone training:

   bash multitask_sh/multitask_distill_cross.sh
   bash sh/average_model.sh

5. Prepare data for the selector:

   bash sh/multitask_distill_cross_inference.sh
   <!-- modify test_split_tag to prepare for different dataset -->

6. UniGenCoder selector training and inference:

   bash multitask_sh/multitask_distill_cross_tune.sh

Citation

If you find this code to be useful for your research, please consider citing:

@article{DBLP:journals/corr/abs-2502-12490,
  author       = {Liangying Shao and
                  Yanfu Yan and
                  Denys Poshyvanyk and
                  Jinsong Su},
  title        = {UniGenCoder: Merging Seq2Seq and Seq2Tree Paradigms for Unified Code
                  Generation},
  journal      = {CoRR},
  volume       = {abs/2502.12490},
  year         = {2025},
  url          = {https://doi.org/10.48550/arXiv.2502.12490},
  doi          = {10.48550/ARXIV.2502.12490},
  eprinttype    = {arXiv},
  eprint       = {2502.12490},
  timestamp    = {Wed, 19 Mar 2025 11:49:46 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2502-12490.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}