card: three-seed checkpoint layout, evaluate.py usage

README.md

@@ -196,6 +196,10 @@ The CSVs are shipped unsplit. The paper's protocol is 80/10/10 train/validation/
 for the multilabel and regression tasks. `train_finetuning.py` reproduces this via `--use_scaffold`,
 `--train_frac`/`--val_frac`/`--test_frac` and `--seed`; reported numbers are means over three seeds.
 
+The exact values used for each task and seed are recorded in
+`models/finetuned/<task>/seed_<n>/finetuning_optimized_config.json`, so you do not have to set
+them by hand.
+
 ---
 
 ## Models
@@ -215,10 +219,26 @@ encoder and the projections are updated. Includes the SELFIES BPE tokenizer.
 
 ### `models/finetuned/<task>/`
 
-One checkpoint per downstream task (`bace`, `bbbp`, `esol`, `freesolv`, `hiv`, `lipo`,
-`pdbbind_full`, `sider`, `tox21`). Each directory holds the task tokenizer plus `model.pt`, a
-checkpoint dict with `backbone` and `head` state dicts, loadable by `predict.py` in the code
-repository.
+Nine downstream tasks (`bace`, `bbbp`, `esol`, `freesolv`, `hiv`, `lipo`, `pdbbind_full`,
+`sider`, `tox21`), each with all three seeds of its multi-seed run:
+
+```
+models/finetuned/<task>/
+├── seed_42/
+├── seed_123/
+└── seed_3407/
+```
+
+Every seed directory is self-contained:
+
+- `model.pt` — a checkpoint dict with `backbone` and `head` state dicts, loadable by `predict.py`
+  and `evaluate.py` in the code repository
+- `finetuning_optimized_config.json` — the full hyperparameter set behind that checkpoint
+  together with the dataset paths and the split settings. Pass it to
+  `train_finetuning.py` to retrain the task from the pretrained backbone, or to `evaluate.py` to
+  score the checkpoint.
+- the task tokenizer and `config.json`
+
 
 ### `models/DMGI/dmgi_model.pt`
 
@@ -277,16 +297,41 @@ has_kg = np.linalg.norm(kg, axis=1) != 0                  # zero vector == modal
 
 ### Fine-tune from the pretrained backbone
 
+Each task ships the hyperparameters that produced its checkpoints, so the recommended way to
+retrain is to hand `train_finetuning.py` that config instead of re-deriving the flags.
+
+The paths inside the configs are relative to the code repository root, with this dataset mounted at
+`data/`. Download into `data/` from the root of your clone and everything lines up:
+
 ```bash
+cd /path/to/SELFormerMM          # the code repository
+
 huggingface-cli download HUBioDataLab/SELFormerMM --repo-type dataset \
-  --include "models/SELFormerMM/*" "finetuning_datasets/classification/bbbp/*" --local-dir ./selformermm
+  --include "models/SELFormerMM/*" "models/finetuned/bbbp/*" \
+            "finetuning_datasets/classification/bbbp/*" --local-dir ./data
 
 python train_finetuning.py \
-  --model_path ./selformermm/models/SELFormerMM \
-  --dataset_meta_csv ./selformermm/finetuning_datasets/classification/bbbp/bbbp.csv \
-  --dataset_embs_npz ./selformermm/finetuning_datasets/classification/bbbp/bbbp_embs.npz \
-  --task_type binary --label_column p_np --num_labels 1 \
-  --use_scaffold 1 --epochs 50 --backbone_lr 1e-5 --head_lr 1e-4 \
+  --config data/models/finetuned/bbbp/seed_42/finetuning_optimized_config.json \
+  --save_dir ./runs/bbbp
+```
+
+`models/SELFormerMM` in the dataset becomes `data/models/SELFormerMM` locally, which is exactly what
+the config's `pretrained_multimodal_dir` points at. If you download somewhere else, override the
+three path flags (`--pretrained_multimodal_dir`, `--dataset_meta_csv`, `--dataset_embs_npz`); any
+flag on the command line wins over the config, which is also how you sweep one setting
+(`--seed 123`, `--head_lr 1e-4`, …) while keeping the rest fixed.
+
+Without a config the equivalent call is:
+
+```bash
+python train_finetuning.py \
+  --model_path HUBioDataLab/SELFormer \
+  --pretrained_multimodal_dir data/models/SELFormerMM \
+  --dataset_meta_csv data/finetuning_datasets/classification/bbbp/bbbp.csv \
+  --dataset_embs_npz data/finetuning_datasets/classification/bbbp/bbbp_embs.npz \
+  --task_type binary --label_column p_np --num_labels 2 \
+  --use_scaffold 1 --seed 42 --epochs 50 \
+  --batch_size 32 --backbone_lr 5e-5 --head_lr 5e-5 --weight_decay 0.01 \
   --save_dir ./runs/bbbp
 ```
 
@@ -294,13 +339,41 @@ python train_finetuning.py \
 
 ```bash
 python predict.py \
-  --model_dir ./selformermm/models/finetuned/bbbp \
-  --input_meta_csv ./selformermm/finetuning_datasets/classification/bbbp/bbbp.csv \
-  --input_embs_npz ./selformermm/finetuning_datasets/classification/bbbp/bbbp_embs.npz \
-  --task_type binary --num_labels 1 --label_column p_np \
+  --model_dir data/models/finetuned/bbbp/seed_42 \
+  --input_meta_csv data/finetuning_datasets/classification/bbbp/bbbp.csv \
+  --input_embs_npz data/finetuning_datasets/classification/bbbp/bbbp_embs.npz \
+  --task_type binary --num_labels 2 --label_column p_np \
   --output_csv ./bbbp_predictions.csv
 ```
 
+### Score the released checkpoints
+
+`predict.py` writes predictions but no metrics. `evaluate.py` rebuilds the split a checkpoint was
+trained with and prints them. Point it at the task directory to score all three seeds and get the
+mean ± standard deviation, which is the form the table above is stated in:
+
+```bash
+python evaluate.py \
+  --model_dir data/models/finetuned/bbbp \
+  --config data/models/finetuned/bbbp/seed_42/finetuning_optimized_config.json \
+  --device cuda
+```
+
+```
+[evaluate] 3 seeds: 42=0.9484  123=0.9264  3407=0.9605
+[evaluate] roc-auc = 0.9451 +/- 0.0141  (mean +/- std over seeds)
+```
+
+Each seed is scored on its own split, taken from the directory name, so one checkpoint cannot
+accidentally be evaluated under another's. For a single seed, pass that seed's config and leave
+`--model_dir` out:
+
+```bash
+python evaluate.py \
+  --config data/models/finetuned/bbbp/seed_123/finetuning_optimized_config.json \
+  --device cuda
+```
+
 ### Pretrain on your own corpus
 
 Supply a SELFIES CSV plus a `.npy` per modality with identical row ordering; substitute a