DeliverBench

Benchmarking AI agents on continuous professional project delivery

🌐 Project website   ·   💻 Code   ·   📦 Download & run

Professional work unfolds through shared files, new inputs, feedback, and revision. DeliverBench evaluates that continuous project lifecycle, from intermediate submissions to a complete and usable final package.

_{The evaluation unit expands from individual tasks to a complete professional project.}

Three public cases

Case	Professional workflow	Stages	Explore
Case 01	Project Mars Refinancing	13	📁 Files · 📦 ZIP
Case 02	Pharmacovigilance ICSR	12	📁 Files · 📦 ZIP
Case 03	StarPulse Legal Delivery	8	📁 Files · 📦 ZIP

Each case also has one independent FINAL evaluation. This public collection contains three complete projects; it is not the entire research collection.

Download and run

Clone the code repository, install its dependencies, and run:

python scripts/download_data.py --project all --output-dir data

The code pins this dataset to a full commit SHA and checks every downloaded project file. Each project is also browsable under projects/; archives/ provides a complete ZIP for each project. SHA256 lists cover the expanded files and archives.

Install shared extension 0.7.2 with stock Harbor 0.20.0 in Python 3.12. Docker and Compose are required. Each project README explains execution and an optional local synthetic pipeline check that does not call paid model APIs. Real Candidate, Judge, and Responder calls require the user's own provider configuration and may incur cost.

How delivery works

_{The same agent carries its work across stages. Checkpoints and the final delivery package are evaluated separately.}

A project in practice

_{Case 01: earlier analyses, models, and lender feedback feed into a six-file final handoff. Original paper artwork.}

Dataset structure

The viewer contains one index row per project, with its directory, domain, stage count, version, and criterion counts. The test split denotes public evaluation tasks; it is not a training split or a set of independent question-answer records. Full documents, staged inputs, rubrics, response rules, authored references, and environment files are in the project directories. Office and PDF attachments retain their original formats.

An agent continues its own session across stages. Later inputs arrive at defined times; response rules generate feedback from its frozen deliverables. The background opening is unscored. The 33 process stages and three FINAL units are scored separately; FINAL does not average process scores or inherit earlier hurdle hits. Incomplete judgments and runtime failures do not become valid zero scores.

Materials and intended use

The tasks reconstruct professional workflows with source documents, reconstructed scenario materials, staged client feedback, rubrics, and authored reference fixtures. Original notices and business-relevant document comments and revisions are retained. The publishers have confirmed permission to redistribute this release's materials.

References, rubrics, simulator internals, and future-stage materials are public for research reproducibility but remain isolated from the Candidate runtime. Evaluated agents must not retrieve these hidden materials from this public repository. These are public evaluation tasks; exposure or training contamination may affect comparisons. The three projects are not representative of every profession or language. Their source material mixes Chinese and English, with output-language requirements defined by the individual tasks. Agent and Judge performance can depend on provider and tool configuration. No universal assessment of professional competence is implied.

Compatibility and licenses

Shared extension: 0.7.2. Harbor: 0.20.0. The Case 01 language protocol treats a prohibited Chinese body as a violation; English body text does not trigger that hurdle. Structured language observations are checked for consistency. All task thresholds, weights, dependencies, and FINAL definitions are fixed in this release.

Dataset content and presentation figures are CC BY 4.0. Executable code and environment/configuration files are Apache-2.0; see LICENSING.md for the boundary. Use CITATION.bib when referencing this release.

This repository contains no model submissions, scores, judgments, evaluation reports, execution logs, or trajectories. The reference fixtures are authored task components.