Dataset Overview
This dataset consists of human gameplay demonstrations collected from direct Super Mario gameplay.
During data collection, a human player manually controlled Mario and selected discrete jump levels (0–6) at each step while successfully completing game episodes. Each sample pairs a structured game-state prompt with the corresponding human-selected jump action.
All data in this dataset was generated entirely in-house: • No external datasets were used • No third-party annotations or proprietary sources were involved • All prompts and action labels were created through direct gameplay and manual annotation
As a result, this dataset is not subject to external licensing restrictions and is released for research and experimental use.
from datasets import load_dataset
dataset = load_dataset("small-lit/overfit_small-aicrowd-mario-training-data", split="train")
# Dict로 변환 (기존 JSONL과 동일한 형식)
data = [dict(example) for example in dataset]
print(data[:2]) # 첫 2개 예시 출력