Update dataset card with benchmark and ablation results

README.md

@@ -66,8 +66,12 @@ A curated subset of the [SuiteSparse Matrix Collection](https://sparse.tamu.edu/
 │       ├── ACTIVSg2000/
 │       │   └── ACTIVSg2000.mtx
 │       └── ...                   # 867 matrix directories
-└── benchmark_results/
-    └── benchmark_gnp_paper.jsonl  # FGMRES benchmark with 6 preconditioners
+├── benchmark_results/
+│   └── benchmark_gnp_paper.jsonl  # FGMRES benchmark: 838 matrices × 6 preconditioners
+└── ablation_results/
+    ├── ablation_edge_features.json     # Study 1: 3 seeds, 500 epochs (original)
+    ├── ablation_edge_features_v2.json  # Study 2: 3 seeds, 500 epochs (pre-Dirichlet fix)
+    └── ablation_edge_features_v3.json  # Study 3: 5 seeds, 1000 epochs (definitive)
 ```
 
 ### manifest.parquet
@@ -86,7 +90,64 @@ Browsable in the HuggingFace Dataset Viewer. Fields:
 
 ### benchmark_results/
 
-JSONL file with FGMRES solver results per matrix for 6 preconditioners: None, Jacobi, Block Jacobi, ILU(0), AMG (AIR), GMRES-Inner. Includes convergence status, iteration count, residual history, and AUC metrics.
+JSONL file with FGMRES solver results for 838/867 matrices (29 timed out/OOM) using 6 preconditioners.
+
+**FGMRES settings** (matching GNP paper): restart=10, max_iters=100, rtol=1e-8.
+
+Per-entry fields: matrix properties (symmetry, diagonal dominance, scaling factor, zero diagonal count), and per-preconditioner results (construction time/status, convergence, iterations, residual history, Iter-AUC, Time-AUC).
+
+#### Benchmark Results Summary
+
+| Preconditioner | Converged | Construction Failed | Not Converged |
+|---|---|---|---|
+| **ILU(0)** | **376 (44.9%)** | 309 (36.9%) | 153 (18.3%) |
+| AMG (AIR) | 206 (24.6%) | 16 (1.9%) | 616 (73.5%) |
+| GMRES-Inner | 107 (12.8%) | 0 (0.0%) | 731 (87.2%) |
+| None | 64 (7.6%) | 0 (0.0%) | 774 (92.4%) |
+| Block Jacobi | 41 (4.9%) | 577 (68.9%) | 220 (26.3%) |
+| Jacobi | 33 (3.9%) | 603 (72.0%) | 202 (24.1%) |
+
+**Iter-AUC win rates** (best preconditioner per matrix):
+ILU 41.1% · GMRES-Inner 33.9% · AMG 24.4% · Rest 0.5%
+
+**Key findings:**
+- 525/838 (62.6%) matrices solved by at least one preconditioner
+- 313/838 (37.4%) matrices unsolved by any classical method
+- 72% of matrices have zero diagonal entries (→ Jacobi family construction failure)
+- ILU dominates small matrices (<5K), AMG becomes competitive at >10K
+- ILU ∪ AMG covers 511/838 (61.0%); 327 matrices (39.0%) have both ILU and AMG failing
+
+**Comparison with GNP paper** (arXiv 2406.00809v3, same 867 matrices):
+
+| Metric | Ours | GNP paper |
+|---|---|---|
+| ILU construction failures | 309 (36.9%) | 348 (40.1%) |
+| AMG construction failures | 16 (1.9%) | 62 (7.2%) |
+| ILU iter-AUC win rate | 41.1% | ~40% |
+| AMG iter-AUC win rate | 24.4% | ~25% |
+| GNP iter-AUC win rate | — | ~25% |
+| GNP construction failures | — | 0 (0.0%) |
+
+### ablation_results/
+
+GCN vs MPNN ablation studies comparing `ContextResGCN` (GCN with Gershgorin-normalized adjacency) against `ContextResMPNN` (explicit edge function with sum aggregation).
+
+**FGMRES settings**: restart=30, max_iters=300, rtol=1e-6. Training on diffusion + advection only, evaluated on diffusion, advection, and graph_laplacian (OOD). Grid sizes: 16, 24, 32 (train) + 48 (OOD).
+
+#### Study 3 Results (definitive, 5 seeds × 1000 epochs)
+
+| Domain | Model | Convergence | Avg Iterations |
+|---|---|---|---|
+| Diffusion | GCN | **800/800 (100%)** | 90.5 |
+| Diffusion | MPNN | **800/800 (100%)** | 85.7 |
+| Advection | GCN | **800/800 (100%)** | 87.2 |
+| Advection | MPNN | **800/800 (100%)** | 83.0 |
+| Graph Laplacian (OOD) | GCN | **400/400 (100%)** | 63.1 |
+| Graph Laplacian (OOD) | MPNN | **0/400 (0%)** | — |
+
+**Training loss** (nearly identical): GCN 0.091 vs MPNN 0.092.
+
+MPNN is competitive on in-distribution domains but catastrophically fails on OOD graph topology (Barabási-Albert graphs with power-law degree distribution). Root cause: MPNN's explicit edge function overfits to training topology; GCN's Gershgorin-normalized adjacency is degree-invariant by construction. See [ADR-04](https://github.com/Csed-dev/MatrixPFN/blob/main/docs/adr/04-gcn-over-mpnn.md) and [ADR-08](https://github.com/Csed-dev/MatrixPFN/blob/main/docs/adr/08-mpnn-generalization-failure.md).
 
 ## Dataset Creation