# Dr.Sparse SpGEMM training pool extension

498 SuiteSparse matrices in the Dr.Sparse harness `.bin` layout (int32 rows, cols, nnz; int32 row_ptr;
int32 col_ind; float32 values; float32 x), split by rows*cols into level1_small (<1e7),
level2_medium (<1e9), level3_large (<1e11) and level4_huge. Built by `level_sparse/extend_pool.py`
(branch v2) from `pool_ext_selector375.txt` (the selector's 400-matrix pool minus the 25 OTF-100
overlaps and the 114 already in the RL pool) and `pool_ext_suitesparse.txt` (a stratified pick from
the SuiteSparse index balancing nnz/row and problem kind). Matrices whose exact intermediate-product
count exceeds 4e9 were skipped (manifest.csv, status too_many_products). None of the OTF-100 test
matrices is included.

    python level_sparse/extend_pool.py --hf-download KinGeorge/Dr.Sparse-RL-pool-ext --out /big/disk/pool_ext --link-into level_sparse