namkoong-lab/mbabench-modeloff

MBABench (ModelOff subset)

MBABench evaluates LLM agents on end-to-end spreadsheet tasks in finance. Instead of answering a question about a table or editing a single formula, an agent has to build a complete, working financial model in Excel. The result is graded against professional standards along three dimensions: Accuracy, Formula, and Format.

This repository contains the ModelOff subset: 38 financial modeling cases from the ModelOff Financial Modeling World Championships (2012–2019). Each case comes with the starting files given to competitors and a reference solution workbook.

• Paper: MBABench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance, NeurIPS 2026 (Evaluations & Datasets Track)
• Website: https://mbabench.org
• Code (agents, harnesses, and LLM judge): https://github.com/namkoong-lab/MBABench-1.0

Dataset structure

tasks.jsonl                    # one line per task (same content as metadata.json, plus task_dir)
metadata.json                  # task metadata keyed by task folder
task_<index>/
  starting_files/              # given to the agent
    *.xlsx                     # starting workbook (35 of 38 tasks)
    *.pdf                      # information pack with the case and its questions (37 of 38 tasks)
  solution_files/
    *.xlsx                     # reference solution workbook (one per task)

Task metadata (tasks.jsonl)

Field	Description
task_dir	Folder holding the task's files, e.g. task_0
index	Position of the task in this release (0–37)
id	Task ID in the full MBABench benchmark (matches the task numbers used in the paper)
task_name	ModelOff case name (season, round, and section)
task_source	Source competition (always modeloff here)
difficulty	Expert difficulty label: VERY-EASY, EASY, MEDIUM, MEDIUM-HARD, or HARD
fmwc_difficulty	Difficulty assigned by FMWC organizers; always null in this subset
old_id	Legacy internal identifier; not needed for evaluation
num_starting_files, num_solution_files	Number of files in each folder

Difficulty distribution

Very Easy	Easy	Medium	Medium-Hard	Hard	Very Hard
1	6	25	4	2	0

Usage

Download the full task folders:

from huggingface_hub import snapshot_download

path = snapshot_download(repo_id="namkoong-lab/mbabench-modeloff", repo_type="dataset")

To run agents on the tasks and grade their workbooks with the MBABench rubric and LLM judge, see the code repository.

Data collection and annotation

• Source: the cases and starting files come from the ModelOff Financial Modeling World Championships (2012–2019), a discontinued competition that was succeeded by the Financial Modeling World Cup. They were obtained from the competition and from a Google Drive collection of ModelOff materials. Some of the reference solutions come from the same Google Drive collection, and the MBABench team contributed a few that it did not include (see Credits).
• Organization: each case was placed in its own task_<index>/ folder with starting_files/ and solution_files/ subfolders.
• Difficulty annotation: the authors rated each task from 1 to 6 on four sub-categories: Scope, General Modeling, Financial Knowledge, and Implementation in Excel. These were weighted 1/3, 1/3, 1/6, and 1/6, and the rounded weighted sum gives the difficulty level. The full criteria are in Appendix B of the paper.

Credits

• Case materials (starting workbooks and PDF information packs): the ModelOff Financial Modeling World Championships (2012–2019), now succeeded by the Financial Modeling World Cup, and the ModelOff collection on Google Drive.
• Reference solutions: the ModelOff collection on Google Drive and all authors of the solution files it contains. Julian Poeltl from the MBABench team contributed a few solutions that were not provided in that collection.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). You may share and adapt it for non-commercial purposes, provided you credit the sources listed under Credits and cite MBABench.

The ModelOff case materials and the solutions from the Google Drive collection remain the copyright of their respective owners.

Croissant metadata

A Croissant 1.1 file with Responsible AI fields is provided in croissant.json.

Citation

@inproceedings{yen2026mbabench,
  title     = {{MBABench}: Evaluating {LLM} Agents on End-to-End Spreadsheet Tasks in Finance},
  author    = {Thomson Yen and Julian Poeltl and Harshith Srinivas Gear and Yilin Meng and Joshua Fan and Adam Shen and Yili Liu and Ali Bauyrzhan and Siri Du and Haoyang Liu and Daniel Guetta and Hongseok Namkoong},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}