Update model card: compact variant

README.md

@@ -29,13 +29,22 @@ model-index:
         metrics:
           - type: roc_auc
             value: 0.786
-            name: ROC-AUC
+            name: ROC-AUC (full model)
           - type: recall
             value: 0.818
-            name: Recall @ threshold 0.37
+            name: Recall @ threshold 0.37 (full model)
           - type: precision
             value: 0.306
-            name: Precision @ threshold 0.37
+            name: Precision @ threshold 0.37 (full model)
+          - type: roc_auc
+            value: 0.779
+            name: ROC-AUC (compact model)
+          - type: recall
+            value: 0.822
+            name: Recall @ threshold 0.37 (compact model)
+          - type: precision
+            value: 0.297
+            name: Precision @ threshold 0.37 (compact model)
 ---
 
 # Flight delay risk: CatBoost pipeline
@@ -45,19 +54,37 @@ Binary classifier that estimates the probability that a US domestic flight arriv
 time only: airline, origin and destination airports, date, scheduled departure hour, distance and
 scheduled flight time.
 
-This repository holds the single artifact used by the
+This repository holds the two model artifacts used by the
 [Flight Delay Prediction & Risk Analysis](https://github.com/Askar-Turanazarov/FlightDelay-Service)
 service (Streamlit UI with SHAP explanations and an FAA-based expected-loss estimate).
-Live demo: https://huggingface.co/spaces/AskarT/flight-delay-risk
+Live demo: https://flight-delay-askar.streamlit.app · project page: https://huggingface.co/spaces/AskarT/flight-delay-risk
 
 ## Files
 
 | File | Size | Contents |
 |---|---|---|
-| `model_pipeline_main.pkl` | 433 MB | `sklearn.pipeline.Pipeline` = `FlightFeatureEngineer` (custom transformer) + `CatBoostClassifier`, saved with `joblib.dump(..., compress=3)` |
+| `model_pipeline_compact.pkl` | 114 MB | **compact variant**, served by the live demo: the same pipeline trained with `max_ctr_complexity=3` |
+| `model_pipeline_main.pkl` | 433 MB | **full variant**, the diploma reference model (`max_ctr_complexity` default = 4) |
+
+Both files are `sklearn.pipeline.Pipeline` = `FlightFeatureEngineer` (custom transformer) +
+`CatBoostClassifier`, saved with `joblib.dump(..., compress=3)`, and share data, split, features
+and every other training parameter. CatBoost stores counter (CTR) tables for combinations of
+categorical features (`ROUTE` ≈ 4 200 values × `DAY` × `MONTH` × ...); limiting the combinations to
+three features cuts the model from 433 MB / ~1.5 GB RAM to 114 MB / ~0.6 GB RAM at a cost of
+0.007 ROC-AUC. The full model was kept unchanged; the compact one was added so that the public demo
+fits the 1 GB free tier of Streamlit Community Cloud.
+
+| | Full model | Compact model |
+|---|---|---|
+| Training script | `src/train_final.py` | `src/train_compact.py` |
+| ROC-AUC (408 907 hold-out flights) | 0.786 | 0.779 |
+| Recall / precision @ 0.37 | 0.818 / 0.306 | 0.822 / 0.297 |
+| Recall / precision @ 0.50 | 0.660 / 0.396 | 0.657 / 0.386 |
+| Pipeline on disk | 433 MB | 114 MB |
+| RAM after `joblib.load` / peak while loading | ~1.5 GB / ~2.8 GB | ~0.6 GB / ~0.9 GB |
 
-The file is large because CatBoost stores counter (CTR) tables for high-cardinality categorical
-features such as `ROUTE` (about 4 200 values) and `DAY`.
+Smaller configurations (`max_ctr_complexity` 2 and 1, 5–52 MB) were also trained and lost
+0.03–0.05 ROC-AUC; the comparison is in `reports/metrics/compact_experiments.txt` of the service repository.
 
 ## Intended use
 
@@ -84,15 +111,16 @@ ORIGIN_AIRPORT, DESTINATION_AIRPORT, ROUTE, IS_WEEKEND, IS_NIGHT`.
 
 CatBoost 1.2.8 on GPU: 2500 iterations, learning rate 0.03, depth 7, `l2_leaf_reg=12`,
 `scale_pos_weight=4.2`, Bernoulli subsample 0.8, early stopping on AUC (100 rounds).
+The compact variant adds `max_ctr_complexity=3`; nothing else differs.
 
 ## Evaluation (408 907 hold-out flights)
 
-| Metric | Value |
-|---|---|
-| ROC-AUC | 0.786 |
-| Recall @ 0.37 (chosen threshold) | 0.818 |
-| Precision @ 0.37 | 0.306 |
-| Recall / Precision @ 0.50 | 0.660 / 0.396 |
+| Metric | Full model | Compact model |
+|---|---|---|
+| ROC-AUC | 0.786 | 0.779 |
+| Recall @ 0.37 (chosen threshold) | 0.818 | 0.822 |
+| Precision @ 0.37 | 0.306 | 0.297 |
+| Recall / Precision @ 0.50 | 0.660 / 0.396 | 0.657 / 0.386 |
 
 The decision threshold 0.37 was chosen to favour recall: a missed delay costs the airline more
 than a false alarm. Raw `predict_proba` is not calibrated.
@@ -114,7 +142,7 @@ import os, sys, joblib, pandas as pd
 from huggingface_hub import hf_hub_download
 
 sys.path.insert(0, os.getcwd())    # repository root, so that `src.feature_engineering` resolves
-path = hf_hub_download("AskarT/flight-delay-catboost", "model_pipeline_main.pkl")
+path = hf_hub_download("AskarT/flight-delay-catboost", "model_pipeline_compact.pkl")  # or model_pipeline_main.pkl
 pipe = joblib.load(path)
 
 flight = pd.DataFrame([{
@@ -130,6 +158,9 @@ print(f"P(delay > 15 min) = {prob:.3f}, flagged = {prob >= 0.37}")
 
 `DEP_HOUR` / `ARR_HOUR` are the scheduled departure and arrival hours (0–23); `IS_WEEKEND` is 1 for
 Saturday and Sunday. The service computes them in `src/app.py` before calling the pipeline.
+The two variants give slightly different probabilities for the same flight (this example: 0.469 compact,
+0.375 full); the service selects the variant with the `MODEL_VARIANT` environment variable
+(`compact` by default, `full` for the diploma model).
 
 ## Limitations