# Genμ Challenge(U&Me Workshop) @ICCV'25 — Data

来源：https://www.kaggle.com/competitions/gen-challenge-u-me-workshop-iccv-25/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

This page describes the data files provided for the unlearning challenge and their formats. In this competition, participants are tasked with generating images that demonstrate successful unlearning of specific concepts (e.g., images in the Apple_Fruit/ folder should not depict apples). Participants will use a provided Google Colab notebook to evaluate their generated images using a fixed Llava model, producing a submission.csv file that contains the model’s responses to these images. This file is then submitted to Kaggle for scoring using the ERB (Evaluation of Retention and Behavior) metric.

The goal is to ensure that the Llava model does not recognize the unlearned concept in the submitted images, meaning the ideal model_response in submission.csv should be “no” for all prompts. A little time spent understanding the data here can save effort later when preparing submissions.

Participants should be able to answer these questions after reading this description:

# What files do I need? 
You need challenge_dataset.csv to run the [Colab](https://colab.research.google.com/drive/1_f6AIjZ5IYeJuVSl27upzSePpTd6-y3U?usp=sharing) notebook and generate submission.csv. The sample_submission.csv file shows the expected submission format.

# What should I expect the data format to be? CSVs with specific columns (detailed below) and a zip file structure for your image submissions.

# What am I predicting? 
You’re not predicting directly; instead, you’re generating images and using the Colab notebook to produce a submission.csv with model responses (“yes” or “no”) indicating whether the Llava model recognizes the unlearned concept.

# What acronyms will I encounter? 
ERB (Evaluation of Retention and Behavior), the metric used for scoring, and Llava (the vision-language model used for evaluation).