Pathfinder-X2
license: CC BY 4.0, Free to use for any purpose, including commercial, with attribution.
The Pathfinder and Pathfinder-X datasets have been crucial for training Large Language Models with Long-Range Dependencies.
In January of 2023, Meta's Mega LLM scored a 97% on the Pathfinder-X dataset, indicating a need for an even more challenging benchmark.
Pathfinder-X2 contains 200,000 512x512 images along with 200,000 segmentation masks for those images. Each image contains an assortment
of dashed-line "snakes" of varying length, and a model's task is to segment only the snake with a circle on one end.
Each image is meant to be fed in as a sequence,pixel-by-pixel, into a Large Language Model.
Explanation paper: https://www.overleaf.com/read/rpsmdnxbdfjt
Based on the Pathfinder dataset by Drew Linsley, Alekh K Ashok, Lakshmi N Govindarajan, Rex Liu, and Thomas Serre.
