This is a repository for the two-round self-improvement SFT training described in the paper "FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation". It is adapted from Llama-Factory. This repository only contains the SFT training code of the FullStack-Learn method, while the Repository Back-Translation and Repository Augmentation code is in the FullStack-Dev repository.
| Model Name | Huggingface Link |
|---|---|
| FullStack-Learn-LM-30B-A3B | 🤗 luzimu/FullStack-Learn-LM-30B-A3B |
Run the following commands:
# install from source
git clone https://github.com/mnluzimu/FullStack-Learn.git
cd FullStack-Learn
conda create -p ./env/fullstack-learn python==3.11 -y
conda activate ./env/fullstack-learn
pip install -e ".[torch,metrics,deepspeed]" --no-build-isolation
# install wandb
pip install wandb
# install megatron
pip install megatron-core
pip install "git+https://github.com/alibaba/roll.git#subdirectory=mcore_adapter"
# install transformer engine (first cd to a path outside this repository)
git clone https://github.com/NVIDIA/TransformerEngine.git
cd TransformerEngine
pip install -U pip wheel setuptools
pip install ninja cmake pybind11
pip install cython
python -m pip install -v --no-build-isolation .
# install flash_attn
pip install flash-attn --no-build-isolationThe config files for the two rounds of training are in examples/megatron. The scripts for running the experiments are in code/train. Set WANDB_API_KEY to your key when launching training tasks to log in with your W&B account.
# First round of training
bash code/train/qwen3_coder_30b_fullstack-learn_round1_2k.sh
# Second round of training
bash code/train/qwen3_coder_30b_fullstack-learn_round2_10k.shExperimental results of the two rounds of training tested on with FullStack-Dev on FullStack-Bench are as follows:
If you find our project helpful, please cite:
@misc{lu2026fullstackagentenhancingagenticfullstack,
title={FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation},
author={Zimu Lu and Houxing Ren and Yunqiao Yang and Ke Wang and Zhuofan Zong and Mingjie Zhan and Hongsheng Li},
year={2026},
eprint={2602.03798},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2602.03798},
}