Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

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

TL;DR

The paper tackles the challenge of adapting instruction-following LLMs to the E-commerce domain by introducing EcomInstruct, a large-scale domain-specific instruction dataset built from open benchmarks and richly structured Chain-of-Task (CoT) tasks. By fine-tuning BLOOMZ into EcomGPT on this data, the authors demonstrate superior zero-shot generalization to unseen E-commerce tasks and datasets, outperforming ChatGPT in cross-domain evaluations. Comprehensive ablations show that CoT tasks—especially those with golden labels—are critical for broad generalization, and that task diversification, more than sheer model size, drives performance gains. The work provides a practical blueprint for constructing vertical-domain instruction data to achieve robust cross-task, cross-dataset generalization in specialized NLP domains.

Abstract

Recently, instruction-following Large Language Models (LLMs) , represented by ChatGPT, have exhibited exceptional performance in general Natural Language Processing (NLP) tasks. However, the unique characteristics of E-commerce data pose significant challenges to general LLMs. An LLM tailored specifically for E-commerce scenarios, possessing robust cross-dataset/task generalization capabilities, is a pressing necessity. To solve this issue, in this work, we proposed the first e-commerce instruction dataset EcomInstruct, with a total of 2.5 million instruction data. EcomInstruct scales up the data size and task diversity by constructing atomic tasks with E-commerce basic data types, such as product information, user reviews. Atomic tasks are defined as intermediate tasks implicitly involved in solving a final task, which we also call Chain-of-Task tasks. We developed EcomGPT with different parameter scales by training the backbone model BLOOMZ with the EcomInstruct. Benefiting from the fundamental semantic understanding capabilities acquired from the Chain-of-Task tasks, EcomGPT exhibits excellent zero-shot generalization capabilities. Extensive experiments and human evaluations demonstrate that EcomGPT outperforms ChatGPT in term of cross-dataset/task generalization on E-commerce tasks.

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

TL;DR

The paper tackles the challenge of adapting instruction-following LLMs to the E-commerce domain by introducing EcomInstruct, a large-scale domain-specific instruction dataset built from open benchmarks and richly structured Chain-of-Task (CoT) tasks. By fine-tuning BLOOMZ into EcomGPT on this data, the authors demonstrate superior zero-shot generalization to unseen E-commerce tasks and datasets, outperforming ChatGPT in cross-domain evaluations. Comprehensive ablations show that CoT tasks—especially those with golden labels—are critical for broad generalization, and that task diversification, more than sheer model size, drives performance gains. The work provides a practical blueprint for constructing vertical-domain instruction data to achieve robust cross-task, cross-dataset generalization in specialized NLP domains.

Abstract

Recently, instruction-following Large Language Models (LLMs) , represented by ChatGPT, have exhibited exceptional performance in general Natural Language Processing (NLP) tasks. However, the unique characteristics of E-commerce data pose significant challenges to general LLMs. An LLM tailored specifically for E-commerce scenarios, possessing robust cross-dataset/task generalization capabilities, is a pressing necessity. To solve this issue, in this work, we proposed the first e-commerce instruction dataset EcomInstruct, with a total of 2.5 million instruction data. EcomInstruct scales up the data size and task diversity by constructing atomic tasks with E-commerce basic data types, such as product information, user reviews. Atomic tasks are defined as intermediate tasks implicitly involved in solving a final task, which we also call Chain-of-Task tasks. We developed EcomGPT with different parameter scales by training the backbone model BLOOMZ with the EcomInstruct. Benefiting from the fundamental semantic understanding capabilities acquired from the Chain-of-Task tasks, EcomGPT exhibits excellent zero-shot generalization capabilities. Extensive experiments and human evaluations demonstrate that EcomGPT outperforms ChatGPT in term of cross-dataset/task generalization on E-commerce tasks.
Paper Structure (25 sections, 5 figures, 11 tables)

This paper contains 25 sections, 5 figures, 11 tables.

Table of Contents

  1. Introduction
  2. EcomInstuct: E-commerce Instruction Tuning Dataset
  3. Overview of the EcomInstuct
  4. Raw Data from Open-Source Benchmarks
  5. Raw Data from Atomic Tasks
  6. Mapping Raw Data to Instruction Data
  7. EcomGPT: Training E-commerce Large Language Model with EcomInstuct
  8. Experiments
  9. Experiment Setup
  10. Baselines
  11. Evaluation Metric.
  12. Dataset Split.
  13. Main Experiments
  14. Ablation Experiments on CoT Tasks
  15. Overall Gain from CoT Tasks
  16. ...and 10 more sections

Figures (5)

  • Figure 1: The complete schema of the atomic tasks.
  • Figure 2: An overview of multi-task instruction tuning of EcomGPT for diverse E-commerce tasks.
  • Figure 3: Human Evaluation results.
  • Figure 4: Scaling experiments on the number of training tasks, with the vertical axis representing the Rouge of the EcomGPT on the unseen dataset.
  • Figure 5: Scaling experiments on the number of training samples per task.