Document the reproduction-branch install and evaluation on main.

README.md

@@ -15,13 +15,13 @@ GENCO optimal-power-flow checkpoints on OPFData ([paper](https://arxiv.org/abs/2
 
 ## Evaluate
 
-Use graphkit branch [genco-paper-repro](https://github.com/gridfm/gridfm-graphkit/tree/genco-paper-repro). We do not guarantee that these checkpoints work on `main`, which is under development.
+Use graphkit branch [genco-paper-repro](https://github.com/gridfm/gridfm-graphkit/tree/genco-paper-repro). Graphkit requires Python 3.10, 3.11, or 3.12. The paper numbers come from this branch. [Evaluate on main](#evaluate-on-main) is the same checkpoints on current `main`.
 
 ```bash
 git clone -b genco-paper-repro https://github.com/gridfm/gridfm-graphkit.git
 cd gridfm-graphkit
 pip install -e .
-pip install "huggingface_hub[cli]"
+pip install huggingface_hub
 TORCH_CUDA_VERSION=$(python -c "import torch; print(torch.__version__ + ('+cpu' if torch.version.cuda is None else ''))")
 pip install torch-scatter -f https://data.pyg.org/whl/torch-${TORCH_CUDA_VERSION}.html
 export MLFLOW_ALLOW_FILE_STORE=true
@@ -52,7 +52,36 @@ gridfm_graphkit evaluate \
   --batch_size 512
 ```
 
-Repeat for the other grids and seeds. Download each dataset into `data/<network>/raw`.
+Repeat for the other grids and seeds. Download each dataset into `data/<network>/raw`. The first `evaluate` writes one graph file per scenario before testing.
+
+## Evaluate on main
+
+The same checkpoints run on current [main](https://github.com/gridfm/gridfm-graphkit/tree/main). This was checked at commit `f104900`. `main` does not ship these YAML files, and it rejects them as config version 0. Use the YAML downloaded with the checkpoint, and pass `--config_conversion auto_inline`. That upgrades the config in memory and leaves the file unchanged. The config still reads splits from `scripts/opfdata/splits/` relative to the repository root.
+
+```bash
+git clone https://github.com/gridfm/gridfm-graphkit.git
+cd gridfm-graphkit
+pip install -e .
+pip install huggingface_hub
+TORCH_CUDA_VERSION=$(python -c "import torch; print(torch.__version__ + ('+cpu' if torch.version.cuda is None else ''))")
+pip install torch-scatter -f https://data.pyg.org/whl/torch-${TORCH_CUDA_VERSION}.html
+export MLFLOW_ALLOW_FILE_STORE=true
+
+mkdir -p data/case118_ieee/raw scripts/opfdata/splits
+hf download gridfm/opfdata_case118_ieee --repo-type dataset --local-dir data/case118_ieee/raw
+hf download gridfm/genco-opfdata --include "case118_ieee/seed42/**" --include "splits/**" --local-dir genco-opfdata
+cp genco-opfdata/splits/*.pt scripts/opfdata/splits/
+
+gridfm_graphkit evaluate \
+  --config_conversion auto_inline \
+  --config genco-opfdata/case118_ieee/seed42/HGNSQ_penalty_11_OPFData_case118_default.yaml \
+  --data_path data \
+  --model_path genco-opfdata/case118_ieee/seed42/best_model_state_dict.pt \
+  --normalizer_stats genco-opfdata/case118_ieee/seed42/normalizer_stats.pt \
+  --batch_size 512
+```
+
+Repeat for the other grids and seeds. On IEEE 14, seed 42, this matched the reproduction-branch evaluation.
 
 ## Files