🧮 ABACUS-Data
Unified Training Data for Image Count Understanding and Count-Faithful Generation
ABACUS-Data is the unified training corpus behind ABACUS, a 3B-parameter vision–language model that jointly performs object counting, crowd counting, referring-expression counting, and count-faithful image generation — with no benchmark-specific tuning. This repository aggregates the eight component shards used across pre-training, task training, and… See the full description on the dataset page: https://huggingface.co/datasets/sauradip/ABACUS-Data.
