README.md

---
license: mit
task_categories:
  - question-answering
language:
  - en
tags:
  - guesstimation
  - numerical-reasoning
  - world-knowledge
  - llm-evaluation
size_categories:
  - n<1K
---

# MARBLES Dataset

MARBLES is a physical guesstimation benchmark designed to probe the world models of large language models (LLMs). It is one of three guesstimation datasets introduced in the associated paper, alongside FUTURE and ELECPRED.

## Dataset Description

This dataset contains 15 guesstimation questions that ask how many small objects (M&Ms, marbles, or quarters) fit inside various common containers (Starbucks cup, shot glass, measuring cup, Altoids tin, or card box). Each question has a ground-truth answer obtained through physical measurement.

### Associated Paper

> Chuang, Y. S., Narendran, S., Harlalka, N., Cheung, A., Gao, S., Suresh, S., ... & Rogers, T. T. (2025). **Probing LLM World Models: Enhancing Guesstimation with Wisdom of Crowds Decoding.** In *Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing* (pp. 4699-4713).

**Paper:** https://aclanthology.org/2025.emnlp-main.234/

## Dataset Structure

### Data Files

- `questions.csv`: Contains all 15 guesstimation questions with ground-truth answers
- `images/`: Contains reference images for each question

### Data Fields

| Field | Description |
|-------|-------------|
| `index_question` | Unique question identifier (1-15) |
| `dataset` | Dataset name (MARBLES) |
| `content` | The guesstimation question in natural language |
| `true_answer` | Ground-truth answer obtained through physical measurement |
| `image` | Filename of the corresponding reference image |

### Example

```
Question: How many standard-sized M&Ms does it take to fill a Starbucks iced tall cup?
Answer: 382
Image: MM_starbucks.png
```

## Objects and Containers

**Objects:**
- Standard-sized M&Ms
- Standard-sized U.S. marbles
- U.S. quarters

**Containers:**
- Starbucks iced tall cup
- Single-shot shot glass
- One cup dry ingredient measuring cup
- Altoids tin container
- Standard-sized Bicycle playing cards box

## Images

While the original paper focuses on text-based guesstimation (without visual input), we include reference images for each question-answer pair. These images show the objects inside the containers and may be useful for multimodal research or verification purposes.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/MARBLES")
```

## Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{chuang-etal-2025-probing,
    title = "Probing {LLM} World Models: Enhancing Guesstimation with Wisdom of Crowds Decoding",
    author = "Chuang, Yun-Shiuan  and
      Narendran, Sameer  and
      Harlalka, Nikunj  and
      Cheung, Alexander  and
      Gao, Sizhe  and
      Suresh, Siddharth  and
      Hu, Junjie  and
      Rogers, Timothy T.",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.234/",
    doi = "10.18653/v1/2025.emnlp-main.234",
    pages = "4699--4713",
    ISBN = "979-8-89176-332-6"}
```

## License

This dataset is released under the MIT License.