Datasets:

ServiceNow-AI
/

gitlab-tasks

license: apache-2.0
task_categories:
  - text-generation
tags:
  - web-agent
  - evaluation
  - gitlab
pretty_name: GitLab Tasks
dataset_info:
  features:
    - name: benchmark
      dtype: string
    - name: task_name
      dtype: string
    - name: task_seed
      dtype: int64
    - name: task_config
      dtype: string
    - name: goal
      dtype: string
    - name: start_url
      dtype: string
    - name: validation
      dtype: string
  splits:
    - name: train
      num_bytes: 806672
      num_examples: 192
  download_size: 57836
  dataset_size: 806672
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

GitLab Tasks

An evaluation dataset for web agents operating on a self-hosted GitLab instance. Each row contains a task goal, a configuration dict, and a standalone Python validation function that scores agent performance by querying the GitLab API.

Dataset overview

155 samples across 23 task categories
Validators are self-contained Python functions that call the GitLab REST API to verify outcomes

Category	Samples	Description
`add-followed-users-to-new-repo`	3	Follow users, create a group, create a project, and add them as collaborators
`create-group`	5	Create a group and add members
`create-issue`	13	Create an issue in a repository
`create-milestone`	5	Create a milestone with start and end dates
`create-pr`	6	Create a merge request between branches
`create-repo`	20	Create a new project with visibility and member settings
`follow-user`	5	Follow one or more users
`fork-repo`	6	Fork a repository
`get-and-update-info`	3	Update status and profile, then report repository contributions
`invite-user`	15	Invite a user to a repository with a specific role
`list-issues`	12	List issues matching specific filters
`list-pr`	2	List merge requests matching specific filters
`list-repos`	6	List repositories matching criteria
`open-issue`	10	Find and open a specific existing issue
`post-message`	8	Post a comment on a merge request
`show-command`	5	Show the git clone command for a repository
`star-fork-create-milestone-issues`	3	Star and fork a repo, create a milestone, and create issues
`star-repo`	5	Star the top N most-starred repositories
`update-issue`	5	Assign an issue to a user
`update-license`	5	Change a repository's license
`update-multiple-issues-single-repo`	3	Add comments to multiple issues in a repository
`update-profile`	5	Update the user's profile homepage URL
`update-status`	5	Set the user's GitLab status

Schema

Column	Type	Description
`benchmark`	`string`	Benchmark name (`gitlab`)
`task_name`	`string`	Task category (e.g. `gitlab.create-issue`)
`task_seed`	`int`	YAML file ID the sample was generated from
`task_config`	`string`	JSON-encoded dict with task-specific parameters used for validation
`goal`	`string`	Natural language instruction shown to the agent
`start_url`	`string`	URL where the agent should begin (empty, defaults to GitLab home)
`validation`	`string`	Python source code of the validator module

Validation

Each validator module exposes a validate(messages, config) function:

def validate(messages: List[Dict[str, str]], config: Dict[str, Any]) -> int:
    """
    Args:
        messages: Conversation history (list of {"role": ..., "content": ...} dicts).
                  The last assistant message is checked for answer-based tasks.
        config:   The task_config dict for this sample.

    Returns:
        1 if the task was completed successfully, 0 otherwise.

    Raises:
        ValueError: If the task configuration is invalid.
    """

Validators call the GitLab REST API to verify outcomes. The following environment variables must be set:

Variable	Description
`GITLAB_BASE_URL`	Base URL of the GitLab instance (e.g. `https://gitlab.example.com`)
`GITLAB_EXTRA_HTTP_HEADERS`	JSON string containing authentication headers (e.g. `{"Authorization": "Bearer <token>"}`)

Usage

import ast
from datasets import load_dataset

ds = load_dataset("servicenow-ai/gitlab-tasks")

sample = ds["train"][0]
config = ast.literal_eval(sample["task_config"])

# Load the validator
exec(sample["validation"], globals())
score = validate(
    messages=[{"role": "assistant", "content": "Done."}],
    config=config,
)

Task categories

Single-action tasks

create-group / create-repo / create-issue / create-milestone / create-pr: Create a GitLab resource with specific attributes. Validated by querying the API for the created resource.
follow-user / fork-repo / star-repo: Social actions on users or repositories. Validated by checking the user's following list, fork list, or starred projects.
invite-user: Invite a user to a project with a specific access level. Validated by checking project membership.
update-issue / update-license / update-profile / update-status: Modify an existing resource. Validated by reading back the resource state from the API.
post-message: Post a comment on a merge request. Validated by fetching the MR's notes.

Information retrieval tasks

list-issues / list-pr / list-repos: The agent must return a URL with the correct query parameters. Validated by checking the URL prefix and query string.
open-issue: Find and navigate to a specific issue. Validated by matching the URL against the API's issue reference.
show-command: Return the SSH clone command for a repository. Validated by comparing against the API's ssh_url_to_repo.

Multi-step tasks

add-followed-users-to-new-repo: Follow users, create a group, create a project, and add collaborators. Validated by checking all sub-tasks via the API.
get-and-update-info: Update status and profile, then report repository contributions. Validated by checking status, profile URL, and message content.
star-fork-create-milestone-issues: Star a repo, fork it, create a milestone, and create issues. Validated by checking all sub-tasks via the API.
update-multiple-issues-single-repo: Add comments to multiple issues. Validated by checking issue notes.