RegretBench measures whether an AI agent protects state before an instructed action destroys it. Each task is a multi-turn conversation with a coding or operations agent working on a real service. A few ordinary requests come first. Then the user asks for a legitimate maintenance action (a cleanup, a migration, a reset, a retention change) that removes data which exists nowhere else, and the agent is scored on whether it kept a recoverable copy first. The request is reasonable and the user wants it done, so refusing is not the goal: the agent should do the work and make it undoable.

Instructions

First, request access to the RegretBench dataset on the Hugging Face Hub. Once you have access, log in with the huggingface_hub CLI and download the full dataset repository (including workspaces):

pip install huggingface-hub
hf auth login
hf download CHATS-Lab/RegretBench --repo-type dataset --local-dir RegretBench
# large seed files are stored xz-compressed; restore them before building any task
(cd RegretBench && find workspaces -name '*.xz' -exec xz -d {} +)

The tasks run with the RegretBench launcher in the BackUpper repository, which builds the container, starts the domain's environment, drives the agent through the turns and runs the judge:

git clone https://github.com/CHATS-lab/BackUpper && cd BackUpper
export RB_TASKS_DIR=$PWD/../RegretBench/workspaces

# check what one task needs, then run it with and without BackUpper
bash regretbench/run_regretbench.sh --task task21-consolidation-delete --harness claude --dry-run
bash regretbench/run_regretbench.sh --task task21-consolidation-delete --harness claude --judge
bash regretbench/run_regretbench.sh --task task21-consolidation-delete --harness claude --backupper on --judge

regretbench/README.md in that repository describes the environments each domain needs (the web apps, the mail server, the cluster image) and the judge.

Dataset Overview

Environments

Environment	What the agent works on	Task Count
Terminal	a shell in a container holding a service's state (files, SQLite, git, registries, CLIs)	65
PostgreSQL-MCP	a PostgreSQL database reached only through an MCP server	50
Playwright-MCP	web apps driven through a browser: ownCloud (20), a Postmill forum (20), GitLab (10)	50
Kubernetes-MCP	a Kubernetes cluster reached through an MCP server	20
Email-MCP	mailboxes on a real mail server reached through email MCP servers	20
Total		205

Task Sources

Every task rests on three real sources:

Regret moment: a reported case in which a person lost the same kind of thing through the same kind of operation
Environment data: the data the container is built from
Agent trajectory: a recorded trajectory in which an agent carries out the same kind of action

Each task.yaml links the three in source, env_source and traj_source, and a short adaptation note says how the task was aligned with its moment.

Regret moments come from three kinds of places:

Category	Source	Task Count
`creation-productivity`	product communities and issue trackers	102
`social-community`	forums, Reddit, Q&A sites	82
`records-media`	news, official records, archives	21

Data Structure

RegretBench/
├── README.md
├── dataset.csv             one row per task
├── taxonomy_pie.png
└── workspaces/
    ├── terminal/task<N>-<name>/
    ├── postgres/task<N>-<name>/
    ├── playwright/task<N>-<name>/
    ├── k8s/task<N>-<name>/
    └── email/task<N>-<name>/

IDs are aligned: ID in dataset.csv, the task<N> prefix of the directory, and id in task.yaml are the same number

Each task directory contains:

File	Description
`task.yaml`	the turns the agent receives, the MCP servers it reaches, the three sources, and the grading context for the judge (`at_risk_state`, `recovery_gap`, `context_echo`)
`Dockerfile`, `files/`	the task container: the agent's workspace and the service state it builds at image build time
`seed.sh`	playwright and email tasks: seeds the task's state into the external web app or mail server
`cluster/`	k8s tasks: the cluster image and its seed

dataset.csv columns:

Column	Description
`ID`	task number
`Tool`	`terminal`, `postgres`, `k8s`, `email`, or `playwright` with the web app in brackets
`Task Name`	the task directory under `workspaces/<environment>/`
`Category`	where the regret moment comes from (see Task Sources)

The source links live in each task's task.yaml: source (the regret moment), env_source (where the data in the container comes from) and traj_source (the recorded agent trajectory).

Terms of Use

By downloading this Dataset, you agree to comply with the following terms of use:

Restrictions: You agree not to use the Dataset in any way that is unlawful or would infringe upon the rights of others.
Acknowledgment: By using the Dataset, you acknowledge that the Dataset may contain data derived from third-party sources, and you agree to abide by any additional terms and conditions that may apply to such third-party data. Some tasks seed their containers with data from third-party public datasets (linked in each task's env_source); that data remains under the license of its source.
Fair Use Declaration: The Dataset may be used for research if it constitutes "fair use" under copyright laws within your jurisdiction. You are responsible for ensuring your use complies with applicable laws.

Derivatives must also include the terms of use above. The task code and metadata are released under the MIT license.

Citation

@article{li2026regretmeasuringbackupawareness,
  title   = {Back Up Before You Regret: Measuring Backup Awareness and Automating Backup in Coding Agents},
  author  = {Xu Li and Simon Yu and Minzhou Pan and Yiyou Sun and Zhaorun Chen and Bo Li and Weiyan Shi},
  journal = {arXiv preprint arXiv:2610.28934},
  year    = {2026}
}

Downloads last month: 14

Total file size:

109 MB

Paper for CHATS-Lab/RegretBench

Back Up Before You Regret: Measuring Backup Awareness and Automating Backup in Coding Agents

Paper • 2610.28934 • Published 2 days ago