| # Frozen JSON Artifacts for `make reproduce-paper` | |
| This directory contains the canonical aggregated JSON files that the paper's | |
| tables and figures are computed from. Bundling these allows reviewers to | |
| regenerate paper-facing artifacts without re-running 40 GPU-hours of experiments. | |
| | File | Source | Used by | | |
| |---|---|---| | |
| | `rd5_fullscale_aggregated.json` | `results/rd5_fullscale/aggregated/aggregated_results.json` | Table 1 (3 strategies), Fig 1 | | |
| | `tabdpt_probing_3seed.json` | `results/phase7/tabdpt_probing/aggregated_3seed.json` | Table 1 TabDPT row, Fig 1 | | |
| | `tabdpt_causal_3seed.json` | `results/phase7/tabdpt_causal/aggregated_3seed.json` | §4.4 TabDPT causal claim | | |
| | `nam_holdout.json` | `results/phase9_nam_holdout/nam_holdout_results.json` | §4.3 NAM holdout | | |
| | `lofo_primary_endpoint.json` | `results/phase8a/nonlinear_probing/lofo_primary_endpoint.json` | Appendix LOFO table | | |
| | `c1c2_baselines.json` | `results/neurips/c1c2_baselines.json` | §4.5 shuffled-label control | | |
| | `scale_10k_multiseed.json` | `results/scale_10k_multiseed/results.json` | Appendix G N=10K scaling | | |
| | `tabpfn25_fullscale_aggregated.json` | `results/rd6/tabpfn25_fullscale/aggregated/*.json` | Table 7 (v2 vs v2.5) | | |
| To regenerate paper figures/tables from these artifacts without GPU: | |
| ```bash | |
| make reproduce-paper | |
| ``` | |