Datasets:

FlyaiLab
/

ecommerce_last_exam

license: mit
tags:
  - benchmark
linked_spaces:
  - FlyaiLab/ecommerce_last_exam_leaderboard
dataset_info:
  features:
    - name: instance_id
      dtype: string
    - name: domain
      dtype: string
    - name: config
      dtype: string
    - name: user_question
      dtype: string
    - name: docker_image
      dtype: string
  splits:
    - name: test
      num_examples: 120
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
  - config_name: travel
    data_files:
      - split: test
        path: data/travel/test-*
  - config_name: e_commerce
    data_files:
      - split: test
        path: data/e_commerce/test-*

E-Commerce Last Exam

A benchmark for evaluating LLM agents on 120 real-world travel planning and e-commerce tool-use tasks. Each task runs in an isolated Docker container with domain-specific CLI tools and SQLite databases. Agents must search, analyze, and produce structured recommendations.

Repository: alibaba-flyai/ecommerce_last_exam
Evaluation CLI: flyai-bench (pip install flyai-bench)
Leaderboard: FlyaiLab/ecommerce_last_exam_leaderboard

Dataset Summary

E-Commerce Last Exam consists of 120 tasks across two configs:

Config	Tasks	Domains
`travel`	77	Hotel booking, transport routing, attraction planning
`e_commerce`	43	Travel gear, food, electronics, lifestyle shopping

Each task provides a natural-language user question and a pre-built Docker image containing the environment (CLI tools, databases, test harness). Agents interact with the environment via tool calls and produce a structured answer scored 0.00 - 1.00.

Data Fields

Field	Type	Description
`instance_id`	string	Unique task identifier (e.g., `attraction_auckland_extreme_sports_415`)
`domain`	string	Task domain (e.g., `attraction`, `hotel`, `transport`, `consume`)
`config`	string	Benchmark config: `travel` or `e_commerce`
`user_question`	string	Natural-language task description
`docker_image`	string	Docker image for the task environment

Usage

from datasets import load_dataset

# Load all 120 tasks
ds = load_dataset("FlyaiLab/ecommerce_last_exam", split="test")

# Load only travel tasks (77)
ds_travel = load_dataset("FlyaiLab/ecommerce_last_exam", "travel", split="test")

# Load only e-commerce tasks (43)
ds_ecom = load_dataset("FlyaiLab/ecommerce_last_exam", "e_commerce", split="test")

Evaluation

pip install flyai-bench

# Run evaluation on travel config
flyai-bench run --dataset-config travel --limit 5 --dry-run
flyai-bench run --dataset-config travel

# Generate report and submit
flyai-bench report
flyai-bench submit --model your-model --provider your-provider

See the GitHub repository for full documentation.

Citation

If you use this benchmark, please cite:

@misc{ecommerce_last_exam,
  title={E-Commerce Last Exam: A Benchmark for LLM Agent Evaluation on Real-World Tool-Use Tasks},
  author={FlyaiLab},
  year={2026},
  url={https://huggingface.co/datasets/FlyaiLab/ecommerce_last_exam}
}