EcomGPT: Instruction-Tuning Large Language Models with Chain-of-Task Tasks for E-commerce

Authors: Yangning Li, Shirong Ma, Xiaobin Wang, Shen Huang, Chengyue Jiang, Hai-Tao Zheng, Pengjun Xie, Fei Huang, Yong Jiang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments and human evaluations demonstrate that Ecom GPT outperforms Chat GPT in term of cross-dataset/task generalization on E-commerce tasks.
Researcher Affiliation Collaboration 1SIGS, Tsinghua University 2Shanghai Tech University 3DAMO Academy, Alibaba Group 4Peng Cheng Laboratory
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code Yes The Ecom GPT will be public at https://github.com/Alibaba-NLP/Ecom GPT.
Open Datasets Yes We manually collected a wide range of E-commerce natural language processing (NLP) datasets from open data sources, such as academic websites and data competition platforms.
Dataset Splits No The Ecom Instruct dataset is divided into two partitions, namely training and testing." The paper specifies training and testing splits but does not explicitly provide details about a separate validation dataset split.
Hardware Specification Yes All experiments are run on 4 NVIDIA A100 SXM4 80GB GPUs.
Software Dependencies No The paper mentions the use of 'Adam W' optimizer and refers to models like 'BLOOMZ' and 'Chat GPT', and a tool 'Alpaca Garbage Collector' (with a URL), but it does not specify explicit version numbers for programming languages, libraries, or frameworks (e.g., Python version, PyTorch/TensorFlow version, CUDA version) needed for replication.
Experiment Setup Yes Adam W (Loshchilov and Hutter 2017) optimizer is employed for model training, with learning rate set of 2e-5 and weight decay of 0. We utilize a cosine learning rate schedule, warming up over 3% of the training steps. The model is fine-tuned with 3 epochs, with the batch size per device set to 4 and the gradient accumulation step set to 8. The maximum sequence length is 1024.