MMCricBench 🏏

Multimodal Cricket Scorecard Benchmark for VQA

This repository contains the dataset for the paper Mind the (Language) Gap: Towards Probing Numerical and Cross-Lingual Limits of LVLMs.

MMCricBench evaluates Large Vision-Language Models (LVLMs) on numerical reasoning, cross-lingual understanding, and multi-image reasoning over semi-structured cricket scorecard images. It includes English and Hindi scorecards; all questions/answers are in English.

---

Overview

• Images: 1,463 synthetic scorecards (PNG)
  • 822 single-image scorecards
  • 641 multi-image scorecards
• QA pairs: 1,500 (English)
• Reasoning categories:
  • C1 – Direct retrieval & simple inference
  • C2 – Basic arithmetic & conditional logic
  • C3 – Multi-step quantitative reasoning (often across images)

---

Files / Splits

We provide two evaluation splits:

• test_single — single-image questions
• test_multi — multi-image questions

If you keep a single JSONL (e.g., test_all.jsonl), use a list for images in every row. Single-image rows should have a one-element list. On the Hub, we expose two test splits.

---

Data Schema

Each row is a JSON object:

Field	Type	Description
id	string	Unique identifier
images	list[string]	Paths to one or more scorecard images
question	string	Question text (English)
answer	string	Ground-truth answer (canonicalized)
category	string (C1/C2/C3)	Reasoning category
subset*	string (single/multi)	Optional convenience field

Example (single-image):

{"id":"english-single-9","images":["English-apr/single_image/1198246_2innings_with_color1.png"],"question":"Which bowler has conceded the most extras?","answer":"Wahab Riaz","category":"C2","subset":"single"}

Loading & Preview

Load from the Hub (two-split layout)

from datasets import load_dataset

# Loads: DatasetDict({'test_single': ..., 'test_multi': ...})
ds = load_dataset("DIALab/MMCricBench")
print(ds)

# Peek a single-image example
ex = ds["test_single"][0]
print(ex["id"])
print(ex["question"], "->", ex["answer"])

# Preview images (each example stores a list of PIL images)
from IPython.display import display
for img in ex["images"]:
    display(img)

Baseline Results (from the paper)

Accuracy (%) on MMCricBench by split and language.

Model	#Params	Single-EN (Avg)	Single-HI (Avg)	Multi-EN (Avg)	Multi-HI (Avg)
SmolVLM	500M	19.2	19.0	11.8	11.6
Qwen2.5VL	3B	40.2	33.3	31.2	22.0
LLaVA-NeXT	7B	28.3	26.6	16.2	14.8
mPLUG-DocOwl2	8B	20.7	19.9	15.2	14.4
Qwen2.5VL	7B	49.1	42.6	37.0	32.2
InternVL-2	8B	29.4	23.4	18.6	18.2
Llama-3.2-V	11B	27.3	24.8	26.2	20.4
GPT-4o	—	57.3	45.1	50.6	43.6

Numbers are exact-match accuracy (higher is better). For C1/C2/C3 breakdowns, see Table 3 (single-image) and Table 5 (multi-image) in the paper.

Contact

For questions or issues, please open a discussion on the dataset page or email Abhirama Subramanyam at penamakuri.1@iitj.ac.in