pioritize BiLSTM model

README.md

@@ -250,8 +250,8 @@ compact BiLSTM over 30-frame feature tensors on a Modal A10 GPU and writes `temp
 73 hidden units, 0.2174 dropout, 0.0004 learning rate, batch size 54, and 73 epochs
 (https://arxiv.org/abs/2411.11548). Evaluation scores every available trained artifact, writes
 per-model metrics into `evaluation.json`, and records the active artifact in `router_selection.json`.
-Baseline wins ties to keep runtime inference lightweight unless the temporal model clearly improves
-the metrics.
+The current selection policy prefers the BiLSTM temporal model when it is available; the baseline is
+kept as a fallback/reference artifact.
 
 Download the selected artifact and its selection file after evaluation, then place them under:
 
@@ -259,9 +259,9 @@ Download the selected artifact and its selection file after evaluation, then pla
 models/exercise_router/active/
 ```
 
-For the baseline this directory should contain `router.joblib`; for the temporal model it should
-contain `router.pt`. In both cases, include `router_selection.json` when present so the local loader
-can choose the intended active artifact even if multiple artifacts exist.
+For the active BiLSTM router this directory should contain `temporal.pt` and
+`router_selection.json`. Keep `router.joblib` only when you want the baseline artifact available for
+comparison or fallback.
 
 To publish and load the router from Hugging Face, use the setup notes in
 [docs/huggingface-router-release.md](docs/huggingface-router-release.md). The draft model card is in

docs/exercise-router-training-report.md

@@ -5,16 +5,16 @@ Generated after the Modal training run on June 11, 2026.
 ## Summary
 
 The exercise router was trained and evaluated for `squat`, `push_up`, `shoulder_press`, and
-`unknown`. The final active artifact is the scikit-learn baseline because it scored slightly higher
-than the BiLSTM temporal model in the artifact selection evaluation.
+`unknown`. The final active artifact is the BiLSTM temporal model. The scikit-learn baseline is
+retained as a reference and fallback artifact.
 
 | Field | Value |
 | --- | --- |
-| Selected model | `baseline.joblib` |
-| Selected artifact | `router.joblib` |
-| Selection rule | Highest accuracy, then unknown rejection rate; baseline wins ties |
-| Local active path | `models/exercise_router/active/router.joblib` |
-| Temporal artifact path | `models/exercise_router/active/temporal.pt` |
+| Selected model | `temporal.pt` |
+| Selected artifact | `temporal.pt` |
+| Selection rule | Prefer BiLSTM temporal when available; baseline falls back when temporal is missing |
+| Local active path | `models/exercise_router/active/temporal.pt` |
+| Baseline artifact path | `models/exercise_router/active/router.joblib` |
 
 ## Data
 
@@ -48,6 +48,29 @@ The BiLSTM hyperparameters follow the Riccio exercise-classification paper:
 | Batch size | 54 |
 | Final training loss | 0.0003 |
 
+## Model Complexity
+
+Counts were checked from the local trained artifacts with `uv run --extra train`.
+
+| Model | Count type | Value |
+| --- | --- | ---: |
+| Baseline | Neural-network-style trainable parameters | 0 |
+| Baseline | Input features | 2,574 |
+| Baseline | Trees | 800 |
+| Baseline | Total tree nodes | 20,778 |
+| Baseline | Split nodes | 9,989 |
+| Baseline | Leaves | 10,789 |
+| Baseline | Approximate learned scalar state | 35,915 |
+| BiLSTM temporal | Trainable parameters | 294,924 |
+| BiLSTM temporal | Input features per frame | 429 |
+| BiLSTM temporal | Hidden units | 73 |
+| BiLSTM temporal | Layers | 1 |
+| BiLSTM temporal | Output classes | 4 |
+
+The baseline is a tree-based `HistGradientBoostingClassifier`, so it does not have trainable tensor
+parameters in the same sense as a neural network. The BiLSTM parameter count includes both LSTM
+directions and the linear classification head.
+
 ## Training Metrics
 
 | Model | Validation accuracy | Unknown rejection rate |
@@ -58,6 +81,8 @@ The BiLSTM hyperparameters follow the Riccio exercise-classification paper:
 ## Selection Evaluation
 
 The final evaluation scored every available trained artifact on the cached router windows.
+The baseline scored slightly higher on this cache, but the active router is BiLSTM so routing uses
+the temporal pose-window sequence directly.
 
 | Model | Artifact | Accuracy | Unknown rejection rate |
 | --- | --- | ---: | ---: |
@@ -112,12 +137,13 @@ uv run modal run scripts/exercise_router_modal.py --stage train-temporal
 uv run modal run scripts/exercise_router_modal.py --stage evaluate
 ```
 
-Download the selected artifacts after evaluation:
+Download the active artifact and selection file after evaluation. Download `router.joblib` too when
+you want to keep the baseline for comparison or fallback:
 
 ```bash
-uv run modal volume get --force pozify-router-models /router.joblib models/exercise_router/active/router.joblib
 uv run modal volume get --force pozify-router-models /temporal.pt models/exercise_router/active/temporal.pt
 uv run modal volume get --force pozify-router-models /router_selection.json models/exercise_router/active/router_selection.json
+uv run modal volume get --force pozify-router-models /router.joblib models/exercise_router/active/router.joblib
 ```
 
 ## Notes

docs/huggingface-router-model-card.md

@@ -23,17 +23,17 @@ Current selected artifact:
 
 ```json
 {
-  "selected_model": "baseline.joblib",
-  "selected_artifact": "router.joblib",
-  "reason": "highest accuracy, then unknown rejection rate; baseline wins ties"
+  "selected_model": "temporal.pt",
+  "selected_artifact": "temporal.pt",
+  "reason": "prefer BiLSTM temporal when available; baseline falls back when temporal is missing"
 }
 ```
 
 Artifacts:
 
-- `router.joblib`: selected scikit-learn baseline artifact.
+- `temporal.pt`: selected PyTorch BiLSTM temporal model trained over 30-frame feature tensors.
 - `router_selection.json`: active artifact selector used by Pozify runtime loading.
-- `temporal.pt`: PyTorch BiLSTM temporal model trained over 30-frame feature tensors.
+- `router.joblib`: scikit-learn baseline artifact kept for comparison and fallback.
 - `training_report.md`: training and evaluation metrics.
 
 ## Intended Use

docs/huggingface-router-release.md

@@ -93,15 +93,15 @@ export POZIFY_ROUTER_DISABLE_HF=1
 
 ## Expected Artifact Selection
 
-The current trained artifacts select the baseline:
+The current active router selects the BiLSTM temporal artifact:
 
 ```json
 {
-  "selected_model": "baseline.joblib",
-  "selected_artifact": "router.joblib",
-  "reason": "highest accuracy, then unknown rejection rate; baseline wins ties"
+  "selected_model": "temporal.pt",
+  "selected_artifact": "temporal.pt",
+  "reason": "prefer BiLSTM temporal when available; baseline falls back when temporal is missing"
 }
 ```
 
-The BiLSTM temporal artifact is uploaded for reproducibility and future selection, but it is not the
-active artifact because the baseline currently scores slightly higher.
+The baseline artifact is still uploaded for comparison and fallback, but runtime routing uses the
+BiLSTM when `router_selection.json` points at `temporal.pt`.

pyproject.toml

@@ -16,6 +16,7 @@ dependencies = [
     # HF Spaces installs gradio[mcp,oauth], which requires pydantic<=2.12.5.
     "pydantic>=2.11.10,<=2.12.5",
     "scikit-learn>=1.4.0",
+    "torch>=2.2.0",
 ]
 
 [project.optional-dependencies]
@@ -24,7 +25,6 @@ dev = [
 ]
 train = [
     "datasets>=2.20.0",
-    "torch>=2.2.0",
 ]
 
 [tool.uv]

scripts/exercise_router_modal.py

@@ -546,7 +546,7 @@ def evaluate() -> dict[str, Any]:
             {
                 "name": "temporal",
                 "source_artifact": temporal_path,
-                "selected_artifact": "router.pt",
+                "selected_artifact": "temporal.pt",
                 **temporal_evaluation,
             }
         )
@@ -560,12 +560,10 @@ def evaluate() -> dict[str, Any]:
     selected = select_router_candidate(candidates)
     if selected["selected_artifact"] == "router.joblib":
         shutil.copyfile(selected["source_artifact"], MODEL_ROOT / "router.joblib")
-    else:
-        shutil.copyfile(selected["source_artifact"], MODEL_ROOT / "router.pt")
     selection = {
         "selected_model": f"{selected['name']}.{ 'joblib' if selected['name'] == 'baseline' else 'pt' }",
         "selected_artifact": selected["selected_artifact"],
-        "reason": "highest accuracy, then unknown rejection rate; baseline wins ties",
+        "reason": "prefer BiLSTM temporal when available; baseline falls back when temporal is missing",
     }
     _write_json(MODEL_ROOT / "router_selection.json", selection)
     result = {

src/pozify/ml/exercise_router_evaluation.py

@@ -72,7 +72,7 @@ def select_router_candidate(candidates: Sequence[Mapping[str, Any]]) -> Mapping[
 
 def router_candidate_sort_key(candidate: Mapping[str, Any]) -> tuple[float, float, int]:
     return (
+        1 if candidate.get("name") == "temporal" else 0,
         float(candidate.get("accuracy", 0.0)),
         float(candidate.get("unknown_rejection_rate", 0.0)),
-        1 if candidate.get("name") == "baseline" else 0,
     )

src/pozify/ml/exercise_router_inference.py

@@ -18,11 +18,11 @@ HF_REPO_ID_ENV = "POZIFY_ROUTER_HF_REPO_ID"
 HF_REVISION_ENV = "POZIFY_ROUTER_HF_REVISION"
 HF_DISABLE_ENV = "POZIFY_ROUTER_DISABLE_HF"
 MODEL_FILENAMES = (
-    "router.joblib",
     "router.pt",
+    "temporal.pt",
+    "router.joblib",
     "model.joblib",
     "baseline.joblib",
-    "temporal.pt",
     "exercise_router.joblib",
 )
 MIN_FINAL_CONFIDENCE = 0.65

tests/test_exercise_classifier.py

@@ -334,17 +334,16 @@ class ExerciseRouterEvaluationTests(unittest.TestCase):
         self.assertEqual(evaluation.confusion_matrix["shoulder_press"]["unknown"], 1)
         self.assertEqual(evaluation.unknown_rejection_rate, 1.0)
 
-    def test_selects_temporal_only_when_metrics_win(self) -> None:
+    def test_prefers_temporal_when_available(self) -> None:
         baseline = {"name": "baseline", "accuracy": 0.91, "unknown_rejection_rate": 0.8}
-        temporal = {"name": "temporal", "accuracy": 0.92, "unknown_rejection_rate": 0.7}
+        temporal = {"name": "temporal", "accuracy": 0.90, "unknown_rejection_rate": 0.7}
 
         self.assertEqual(select_router_candidate([baseline, temporal]), temporal)
 
-    def test_selects_baseline_on_metric_tie(self) -> None:
+    def test_selects_baseline_when_temporal_is_missing(self) -> None:
         baseline = {"name": "baseline", "accuracy": 0.91, "unknown_rejection_rate": 0.8}
-        temporal = {"name": "temporal", "accuracy": 0.91, "unknown_rejection_rate": 0.8}
 
-        self.assertEqual(select_router_candidate([baseline, temporal]), baseline)
+        self.assertEqual(select_router_candidate([baseline]), baseline)
 
 
 if __name__ == "__main__":

uv.lock

@@ -2191,6 +2191,7 @@ dependencies = [
     { name = "opencv-python-headless" },
     { name = "pydantic" },
     { name = "scikit-learn" },
+    { name = "torch" },
 ]
 
 [package.optional-dependencies]
@@ -2199,7 +2200,6 @@ dev = [
 ]
 train = [
     { name = "datasets" },
-    { name = "torch" },
 ]
 
 [package.metadata]
@@ -2216,7 +2216,7 @@ requires-dist = [
     { name = "pydantic", specifier = ">=2.11.10,<=2.12.5" },
     { name = "ruff", marker = "extra == 'dev'", specifier = ">=0.5.0" },
     { name = "scikit-learn", specifier = ">=1.4.0" },
-    { name = "torch", marker = "extra == 'train'", specifier = ">=2.2.0" },
+    { name = "torch", specifier = ">=2.2.0" },
 ]
 provides-extras = ["dev", "train"]