LoLBench

Overview

Software development often begins with user intent and high-level design rather than an implementation-ready specification. Developers must ground those goals in an existing architecture, identify the affected modules and interfaces, and produce a correct implementation without regressing the rest of the system. LoLBench (Long-horizon Tasks on Large Software Systems) evaluates coding agents in this intent-to-implementation setting.

These tasks combine two forms of difficulty. Perception complexity is the challenge of interpreting an abstract proposal, understanding cross-module dependencies, and locating the code that must change. Implementation complexity is the challenge of turning that understanding into functionally correct, regression-free changes in a large existing system.

LoLBench contains 100 modular development tasks from human-written enhancement proposals across 29 large open-source projects and five domains.

Quick Start

Prerequisites

• Docker >= 20.0
• Python >= 3.9
• Harbor (recommended for agent evaluation)

1. Download Dataset

pip install -U huggingface-hub
hf download lolbench26/LoLBench --repo-type dataset --local-dir data/

Pre-built Docker images are referenced by each task's task.toml and are pulled automatically when using Harbor; no local image build is required.

2. Run Agent Evaluation (Recommended: Harbor)

Install Harbor:

uv tool install harbor

Run a task with Codex:

export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://your-api-endpoint/v1"

harbor run \
    --path data/cpython_1 \
    --agent codex \
    --model "your-model-name" \
    --n-attempts 1 \
    --n-concurrent 1 \
    --yes \
    --jobs-dir results/codex \
    --ae OPENAI_API_KEY="$OPENAI_API_KEY" \
    --ae OPENAI_BASE_URL="$OPENAI_BASE_URL"

Run a task with OpenCode:

export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://your-api-endpoint/v1"

harbor run \
    --path data/cpython_1 \
    --agent opencode \
    --model "your-model-name" \
    --n-attempts 1 \
    --n-concurrent 1 \
    --yes \
    --jobs-dir results/opencode \
    --ae OPENAI_API_KEY="$OPENAI_API_KEY" \
    --ae OPENAI_BASE_URL="$OPENAI_BASE_URL"

Dataset viewer schema

Column	Description
task_id	Harbor task identifier.
language	Project language from the paper's project classification.
domain	Project domain from the paper's project classification.
instruction	Complete contents of the task's instruction.md.
f2p	Overall fail-to-pass suite as test selectors only.
p2p	Overall pass-to-pass suite as test selectors only.
solution	Complete contents of solution/solution.patch.

Task Structure

Each task directory contains the Harbor task definition and its associated assets:

<task_id>/
├── environment/
├── solution/
│   └── solution.patch
├── tests/
├── instruction.md
├── README.md
└── task.toml

The task metadata records the upstream repository, base commit, target merge/solution commit, and published container image. The canonical verifier mode is union, which evaluates the overall original-plus-augmented test suites represented in the viewer.

Languages & Repositories

Language	Repositories	Tasks
C	PHP	4
C#	ASP.NET Core, Roslyn	2
C++	Apache Arrow, DuckDB	3
Go	cert-manager, gRPC-Go, Kueue, Kubernetes, OpenTofu, Prometheus, Vitess	18
Java	Apache Flink, Apache Iceberg, Apache Kafka, OpenJDK, Presto	36
Python	CPython, Django, FastAPI, Mypy, NumPy, Pandas, Scikit-learn	25
Rust	Apache DataFusion, Cargo, Ruff, Rust	8
TypeScript	TypeScript	4

LoLBench-authored benchmark metadata and tooling are released under MIT. Individual upstream projects and incorporated materials remain subject to their respective licenses.