Retrain model from records csv

README.md

@@ -10,10 +10,11 @@ from the order time.
 
 ## Data Assumption
 
-`timeout_records.csv` in the Dataset repository only stores timeout orders, so
-the first model artifact is a positive-only time-risk model. After users submit
-both timeout and non-timeout records through the Space, retraining will switch to
-`labeled_empirical_bayes`, which is closer to a calibrated probability model.
+`records.csv` in the Dataset repository stores the manually maintained base
+records. When both timeout and non-timeout samples are present, retraining uses
+`labeled_empirical_bayes`, which is closer to a calibrated probability model. If
+only timeout samples are available, the artifact remains a positive-only
+time-risk model.
 
 ## Train
 

timeout_model.json

@@ -1,6 +1,6 @@
 {
   "base_probability": 0.35,
-  "generated_at": "2026-05-23T18:07:51.470272+00:00",
+  "generated_at": "2026-05-25T08:33:53.609270+00:00",
   "hour_probabilities": {
     "0": {
       "positive_count": 0,
@@ -14,53 +14,53 @@
     },
     "10": {
       "positive_count": 0,
-      "probability": 0.1391,
-      "smoothed_density": 0.028
+      "probability": 0.1357,
+      "smoothed_density": 0.0231
     },
     "11": {
       "positive_count": 0,
-      "probability": 0.1963,
-      "smoothed_density": 0.1121
+      "probability": 0.1828,
+      "smoothed_density": 0.0923
     },
     "12": {
       "positive_count": 1,
-      "probability": 0.3043,
-      "smoothed_density": 0.271
+      "probability": 0.3031,
+      "smoothed_density": 0.2692
     },
     "13": {
       "positive_count": 1,
-      "probability": 0.3043,
-      "smoothed_density": 0.271
+      "probability": 0.3658,
+      "smoothed_density": 0.3615
     },
     "14": {
-      "positive_count": 0,
-      "probability": 0.1963,
-      "smoothed_density": 0.1121
+      "positive_count": 2,
+      "probability": 0.392,
+      "smoothed_density": 0.4
     },
     "15": {
       "positive_count": 0,
-      "probability": 0.2153,
-      "smoothed_density": 0.1402
+      "probability": 0.3083,
+      "smoothed_density": 0.2769
     },
     "16": {
       "positive_count": 0,
-      "probability": 0.406,
-      "smoothed_density": 0.4206
+      "probability": 0.4338,
+      "smoothed_density": 0.4615
     },
     "17": {
-      "positive_count": 4,
+      "positive_count": 5,
       "probability": 0.8,
       "smoothed_density": 1.0
     },
     "18": {
       "positive_count": 3,
-      "probability": 0.7301,
-      "smoothed_density": 0.8972
+      "probability": 0.732,
+      "smoothed_density": 0.9
     },
     "19": {
-      "positive_count": 0,
-      "probability": 0.3679,
-      "smoothed_density": 0.3645
+      "positive_count": 1,
+      "probability": 0.4914,
+      "smoothed_density": 0.5462
     },
     "2": {
       "positive_count": 0,
@@ -68,19 +68,19 @@
       "smoothed_density": 0.0
     },
     "20": {
-      "positive_count": 0,
-      "probability": 0.1772,
-      "smoothed_density": 0.0841
+      "positive_count": 1,
+      "probability": 0.3188,
+      "smoothed_density": 0.2923
     },
     "21": {
       "positive_count": 0,
-      "probability": 0.12,
-      "smoothed_density": 0.0
+      "probability": 0.1828,
+      "smoothed_density": 0.0923
     },
     "22": {
       "positive_count": 0,
-      "probability": 0.12,
-      "smoothed_density": 0.0
+      "probability": 0.1357,
+      "smoothed_density": 0.0231
     },
     "23": {
       "positive_count": 0,
@@ -132,19 +132,19 @@
       "optional product name"
     ],
     "non_timeout_records": 0,
-    "timeout_records": 9,
-    "training_records": 9
+    "timeout_records": 14,
+    "training_records": 14
   },
   "mode": "positive_only_time_density",
   "model_version": 1,
   "product_probabilities": {},
   "time_bucket_probabilities": {
     "afternoon": {
-      "positive_count": 0,
-      "probability": 0.16
+      "positive_count": 2,
+      "probability": 0.28
     },
     "dinner": {
-      "positive_count": 7,
+      "positive_count": 9,
       "probability": 0.7
     },
     "late_night": {
@@ -153,21 +153,21 @@
     },
     "lunch": {
       "positive_count": 2,
-      "probability": 0.3143
+      "probability": 0.28
     },
     "morning": {
       "positive_count": 0,
       "probability": 0.16
     },
     "night": {
-      "positive_count": 0,
-      "probability": 0.16
+      "positive_count": 1,
+      "probability": 0.22
     }
   },
   "weekday_probabilities": {
     "0": {
-      "positive_count": 0,
-      "probability": 0.18
+      "positive_count": 2,
+      "probability": 0.4
     },
     "1": {
       "positive_count": 0,
@@ -178,8 +178,8 @@
       "probability": 0.29
     },
     "3": {
-      "positive_count": 1,
-      "probability": 0.29
+      "positive_count": 2,
+      "probability": 0.4
     },
     "4": {
       "positive_count": 4,
@@ -190,8 +190,8 @@
       "probability": 0.51
     },
     "6": {
-      "positive_count": 0,
-      "probability": 0.18
+      "positive_count": 2,
+      "probability": 0.4
     }
   }
 }

timeout_risk_model.py

@@ -22,6 +22,7 @@ MODEL_VERSION = 1
 CSV_FIELDS = [
     "product_name",
     "order_time",
+    "timeout",
     "is_timeout",
     "expected_latest_time",
     "actual_arrival_time",
@@ -133,7 +134,11 @@ def read_csv_records(path: Path, forced_timeout: bool | None = None) -> list[Tra
 
 def load_dataset_records(dataset_dir: str | Path) -> list[TrainingRecord]:
     dataset_path = Path(dataset_dir)
-    records = read_csv_records(dataset_path / "timeout_records.csv", forced_timeout=True)
+    records_path = dataset_path / "records.csv"
+    if records_path.exists():
+        records = read_csv_records(records_path, forced_timeout=None)
+    else:
+        records = read_csv_records(dataset_path / "timeout_records.csv", forced_timeout=True)
 
     tmp_dir = dataset_path / "tmp"
     if tmp_dir.exists():

train.py

@@ -25,7 +25,7 @@ def download_dataset_snapshot(repo_id: str) -> Path:
         repo_id=repo_id,
         repo_type="dataset",
         local_dir=str(target_dir),
-        allow_patterns=["timeout_records.csv", "tmp/*.csv"],
+        allow_patterns=["records.csv", "timeout_records.csv", "tmp/*.csv"],
     )
     return target_dir