# Pre-trained model checkpoints

Each subdirectory holds the **minimum artifacts** needed to run the matching script in `evaluation/`, plus metadata for reproducibility. Training-time performance diagnostics (ROC/PR curves, confusion matrices, Optuna trial tables, etc.) are **not** included — regenerate them with `--save-diagnostics` when re-training.

## `nn_one_hot/`

| File | Purpose |
|------|---------|
| `model_weights/one_hot_model_full_model.h5` | Primary Keras checkpoint |
| `model_weights/one_hot_model_weights.h5` | Weights-only fallback |
| `model_weights/one_hot_model_architecture.json` | Architecture fallback (with weights H5) |
| `one_hot_nn_validation_temperature.csv` | Temperature scaling for inference |
| `one_hot_nn_validation_validation_diagnostics_summary.csv` | F1-optimal classification threshold |
| `model_parameters.json` | Training hyperparameters and `target_length` |
| `best_hyperparameters.csv` | Optuna best trial (architecture fallback) |
| `random_seed.txt` | Random seed used during training |

Evaluator: `evaluation/evaluate_nn_one_hot.py`

## `rf_one_hot/`

| File | Purpose |
|------|---------|
| `one_hot_rf_model.joblib` | sklearn RandomForest + embedded `target_length` |
| `one_hot_rf_validation_temperature.csv` | Temperature scaling |
| `one_hot_rf_validation_validation_diagnostics_summary.csv` | F1-optimal threshold |
| `model_parameters.json` | Training metadata |
| `best_hyperparameters.csv` | Optuna best trial |
| `random_seed.txt` | Random seed |

Evaluator: `evaluation/evaluate_rf_one_hot.py`

## `nn_mean_pertoken_esm/`

Per head (`mean` and `per_token`):

| File pattern | Purpose |
|--------------|---------|
| `{head}_embeddings_model_weights.h5` | Keras softmax head |
| `{head}_embeddings_scaler.npy` | `StandardScaler` used at training |
| `{head}_embeddings_pca.npy` | PCA for per-token head only |
| `{head}_embeddings_validation_temperature.csv` | Temperature scaling |
| `{head}_embeddings_validation_validation_diagnostics_summary.csv` | F1-optimal threshold |
| `{head}_embeddings_hyperparameters.csv` | Learning rate, L1, epochs |

Plus `model_parameters.json` and `random_seed.txt`.

Evaluator: `evaluation/evaluate_nn_mean_pertoken_esm.py` — also requires `.pt` embedding caches (not shipped; see `extract_embeddings/`).

## `rf_mean_pertoken_esm/`

Per head (`mean` and `per_token`):

| File pattern | Purpose |
|--------------|---------|
| `{head}_embeddings_rf_model.joblib` | sklearn RandomForest |
| `{head}_embeddings_scaler.joblib` | Fitted scaler |
| `{head}_embeddings_pca.joblib` | PCA for per-token head only |
| `{head}_embeddings_validation_temperature.csv` | Temperature scaling |
| `{head}_embeddings_validation_validation_diagnostics_summary.csv` | F1-optimal threshold |
| `{head}_embeddings_hyperparameters.csv` | RF hyperparameters from Optuna |

Plus `model_parameters.json` and `random_seed.txt`.

Evaluator: `evaluation/evaluate_rf_mean_pertoken_esm.py` — also requires `.pt` embedding caches.

## Composition logistic regression

No bundle is shipped here. Train with `models/lr_sequence_composition_baseline.py`, then point `evaluation/evaluate_lr_sequence_composition.py` at the output directory (needs `composition_lr_model.joblib` + calibration CSVs).