PinNet: Pinpoint Instructive Information for Retrieval Augmented Code-to-Text Generation

Authors: Han Fu, Jian Tan, Pinhan Zhang, Feifei Li, Jianling Sun

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using extensive experiments on code summarization and SQL-to-text generation, we demonstrate that the proposed method can significantly outperform all of the baselines.
Researcher Affiliation Collaboration 1Alibaba Group, Hangzhou, China 2College of Computer Science and Technology, Zhejiang University, Hangzhou, China.
Pseudocode No The paper describes algorithms and components but does not provide structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide a direct link or explicit statement about the availability of its own open-source code.
Open Datasets Yes For code summarization task, we use a widely adopted benchmark, Code XGLUE (Lu et al., 2021)... For SQL-to-text generation task, we conduct experiments on two datasets, Wiki SQL (Zhong et al., 2017) and Stack Overflow (Iyer et al., 2016).
Dataset Splits Yes Table 9. Statistics of training/developing/testing samples and retrieval databases of different datasets. Code XGLEU-Java Train 164,923 Dev 5,183 Test 10,955.
Hardware Specification Yes The models are trained on a single Nvidia A100 GPU.
Software Dependencies No The paper mentions pre-trained models like Graph Code BERT and RoBERTa, and optimizers like Adam, but does not provide specific version numbers for software dependencies such as deep learning frameworks or libraries.
Experiment Setup Yes For both Code XGLEU-Java and Code XGLEU-Python datasets, we truncate the input and output sequence to maximum 512 and 80 tokens, respectively. ... The Pin Net-Ret model is trained with Adam (Kingma & Ba, 2014) using batch size 128. The learning rate is 2 10 5. To train the generation model sufficiently, we set the learning rates of Pin Net-Enc and Pin Net-Dec as 1 10 5 and 1 10 4 , respectively. The generation model is trained with batch size 64. For the retrieval task, we set m1 and m2 to 0.2 and 0.4, respectively. The hyper-parameters of the generation loss, α and β, are both set to 1.0. During training, we use Pin Net-Ret to select the top-10 code summaries by cosine score. At runtime, we use beam search and the beam size is set to 4.