Datasets:

tencent
/

workbuddy-bench

license: other
license_name: tencent
language:
  - en
  - zh
task_categories:
  - text-generation
tags:
  - code
  - agent
  - benchmark
  - swe
  - security
pretty_name: Tencent WorkBuddy Bench
size_categories:
  - n<1K

WorkBuddy Bench — Datasets

The task datasets for WorkBuddy Bench, a benchmark for evaluating coding agents on real-world developer, PM, algo, QA, ops, and security work. This repository hosts only the task data; the evaluation framework, setup, and usage all live in the GitHub repository above.

Dataset Summary

WorkBuddy Bench is built from real-world work tasks. Given a task and a sandboxed workspace, an agent must produce the correct change — a patch, an artifact, or a report — which is then graded against a test suite. The benchmark ships four independent subsets, one .tar.gz archive each:

Subset	Archive	Tasks	Domain
Code	`wb-bench-code-v1.0.tar.gz`	80	repo-level coding tasks
Web	`wb-bench-web-v1.0.tar.gz`	70	front-end / GUI (HTML/CSS/JS)
Office	`wb-bench-office-v1.0.tar.gz`	50	office data / file workflows
Security	`wb-bench-sec-v1.0.tar.gz`	60	security / vuln (red + blue)

SHA256SUMS lists the checksum of every archive.

Dataset Structure

Each archive extracts to wb-bench-<subset>-v1.0/, with one directory per task:

wb-bench-<subset>-v1.0/
  dataset.toml
  tasks/<task-id>/
    instruction.md                 # the task prompt
    task.toml                      # task metadata
    environment/                   # Dockerfile, docker-compose, scorer, workspace.tar.gz
    tests/                         # grading suite

Each task's environment/workspace.tar.gz stays packed; the task Dockerfile extracts it into /workspace inside the container.

Usage

These archives are consumed by the WorkBuddy Bench framework, not loaded directly. See the GitHub repository for how to download, extract, and run evaluations.