This dataset contains a collection of totally unimodular (TU) node–arc incidence matrices, generated from random directed graphs with node counts between 50,000 and 100,000. Each column of the incidence matrix has exactly one +1 (arc tail) and one –1 (arc head). Because of the TU property, all linear programming relaxations of integer flow problems are guaranteed to have integer solutions. The dataset includes: matrices.csv: sparse representation of all instances (two rows per arc, +1 and –1). metadata.csv: summary of each instance (nodes, arcs, density). Conversion scripts (make_dat_all.py, make_dat_all.R) to produce .dat files for AMPL or other solvers. Typical applications include benchmarking very large-scale network flow and minimum-cost flow solvers, and exploring algorithmic scalability at extreme graph sizes. ⚠️ Large file notice: The CSV files in this collection are extremely large (≈25 GB each). They cannot usually be opened directly in spreadsheet software or loaded fully into memory on a typical laptop or desktop. For analysis, we recommend: Chunked reading (e.g. pandas.read_csv(..., chunksize=...)), Out-of-core frameworks such as Dask or Polars, or Importing into a database (e.g. PostgreSQL, SQLite). For smaller and more manageable datasets, please see the Small, Medium A, and Medium B collections.