Upload folder using huggingface_hub

.claude/session_2026-05-07_fastapi_setup.md

@@ -427,3 +427,204 @@ Côté Kevin (manuel, une seule fois) :
 3. Vérifier la présence du parquet sur HF Dataset.
 4. Re-trigger le `workflow_dispatch` du CI → build Docker passe, Space
    démarre, lifespan télécharge le parquet une fois et le cache.
+
+---
+
+## Session 2026-05-08 — Code review complète + Tier 1 cleanup
+
+Quatre subagents lancés en parallèle (python-reviewer, security-reviewer,
+code-reviewer, architect) sur l'ensemble du repo. Tests post-cleanup :
+**45/45 pass, ruff clean, coverage 96.56 %**.
+
+### Fix critique de la session : libgomp.so.1 sur HF Space
+
+`OSError: libgomp.so.1: cannot open shared object file` au cold start.
+LightGBM dépend d'OpenMP au runtime, absent de `python:3.12-slim`.
+
+**Faux ami `packages.txt`** : ce fichier n'est lu par HF Spaces que pour
+`sdk: gradio` ou `sdk: streamlit`. Pour `sdk: docker`, **il est ignoré** —
+les paquets système doivent être installés dans le `Dockerfile` via
+`apt-get`.
+
+Fix appliqué (Dockerfile) :
+
+```dockerfile
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends libgomp1 \
+    && rm -rf /var/lib/apt/lists/*
+```
+
+`packages.txt` supprimé du repo.
+
+### Tier 1 — fixes appliqués (12 items, zéro changement comportemental)
+
+| Fichier | Fix |
+|---------|-----|
+| `api/inference_assembler.py:93` | drop `inplace=True` après `.copy()` |
+| `api/main.py:115-117` | comment FR retiré, return type, tag `meta` |
+| `api/predictor.py` + `api/schemas.py` | `Decision` dédupliqué (DRY) |
+| `api/schemas.py` | `Optional[X]` → `X \| None` (PEP 604, Pydantic v2) |
+| `feature_engineering/aggregations.py:201-202` | `.apply(lambda)` → `.clip(lower=0)` |
+| `feature_engineering/orchestrator.py` | `print()` → `logger.info()` |
+| `scripts/build_no_history_template.py` | `list(...)`, type widening `\| None` |
+| `scripts/export_model.py:55` | guard `if mv.run_id is None` (mypy) |
+| `scripts/export_model.py` | emojis `✅⚠️` → `[OK]`/`[WARN]` |
+| `api/settings.py:42-44` | comment expliquant le seuil 0.33 |
+| `README.md` | XGBoost → LightGBM (5 occurrences + badge) |
+| `README.md` | exemple `decision: "GRANTED"` au lieu de `false` |
+
+### Tier 2 — suggestions reportées (à attaquer plus tard si besoin)
+
+#### A. Split deps prod / offline → image Docker -700 MB à -1 GB
+
+`pyproject.toml` actuellement liste comme deps prod : `mlflow`, `optuna`,
+`shap`, `xgboost`, `jupyter`, `ipykernel`, `matplotlib`, `seaborn`,
+`plotly`. Aucune n'est importée par `api/`. Plan :
+
+```toml
+[project]
+dependencies = [
+    "fastapi>=0.136.1",
+    "huggingface-hub>=1.14.0",
+    "joblib",            # ajouter explicitement
+    "lightgbm>=4.0.0",
+    "numpy>=2.4.3",
+    "pandas==2.3.3",
+    "pyarrow>=23.0.1",
+    "pydantic>=2.13.3",
+    "uvicorn[standard]>=0.46.0",
+]
+
+[dependency-groups]
+dev = ["httpx", "pytest", "pytest-cov", "ruff", "mypy"]
+offline = [
+    "mlflow>=2.18,<3", "optuna", "shap", "xgboost", "matplotlib",
+    "seaborn", "plotly", "jupyter", "ipykernel",
+]
+```
+
+Dockerfile reste sur `uv sync --frozen --no-dev` (le groupe offline est
+exclu par défaut sauf si activé).
+
+**Risque** : faible. Vérifier que `api/predictor.py` n'importe pas mlflow
+au runtime (à confirmer — le modèle est désérialisé via joblib mais
+`joblib.load` peut tirer mlflow lors de l'unpickle si le modèle est un
+`PyFuncModel`). Si oui → alternative : option B ci-dessous.
+
+#### B. Re-export modèle en LightGBM Booster natif → 3-5× faster + supprime mlflow
+
+Actuellement `models/model.joblib` est un `mlflow.pyfunc.PyFuncModel`
+contenant le Booster. À l'inférence on appelle `model.predict(df)` qui
+re-route via le wrapper PyFunc.
+
+Plan :
+
+```python
+# Dans OC_P6, à côté de la sauvegarde MLflow :
+booster = lgbm_model.booster_   # ou .estimator.booster_ si pipeline
+booster.save_model("models/model.txt")
+```
+
+Côté API (`predictor.py`) :
+
+```python
+import lightgbm as lgb
+model = lgb.Booster(model_file=str(model_path))
+proba = model.predict(features.values)[0]
+```
+
+**Bénéfice** : élimine mlflow du runtime (déblocage du A), latence
+3-5× plus faible sur single-row, image plus légère.
+
+**Risque** : moyen. Demande un test de parité (probabilité identique
+±1e-9 entre l'ancien et le nouveau modèle sur un échantillon de 1000
+clients).
+
+#### C. Multi-stage Dockerfile + pin base image
+
+```dockerfile
+FROM python:3.12.10-slim AS builder
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential && rm -rf /var/lib/apt/lists/*
+COPY --from=ghcr.io/astral-sh/uv:0.5.4 /uv /usr/local/bin/uv
+WORKDIR /app
+COPY pyproject.toml uv.lock ./
+RUN uv sync --frozen --no-dev --no-install-project
+
+FROM python:3.12.10-slim AS runtime
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgomp1 && rm -rf /var/lib/apt/lists/*
+COPY --from=builder /opt/venv /opt/venv
+ENV PATH=/opt/venv/bin:$PATH
+WORKDIR /app
+COPY api/ ./api/
+COPY models/ ./models/
+EXPOSE 7860
+CMD ["uvicorn", "api.main:app", "--host", "0.0.0.0", "--port", "7860"]
+```
+
+**Bénéfice** : -100-200 MB sur l'image finale (pas de cache uv,
+pas de build-essential), reproductibilité (digest pinning possible
+en remplaçant `:3.12.10-slim` par `@sha256:...`).
+
+**Risque** : minimal.
+
+#### D. DataFrame downcast au cold start → RAM -40-60 %
+
+Le parquet (235 MB sur disque) → ~1.0-1.8 GB en RAM avec dtypes par défaut
+(float64, int64). Conversion :
+
+```python
+# Dans InferenceArtefacts.load() :
+feature_store = pd.read_parquet(feature_store_path)
+for col in feature_store.select_dtypes(include="float64").columns:
+    feature_store[col] = feature_store[col].astype("float32")
+for col in feature_store.select_dtypes(include="int64").columns:
+    feature_store[col] = pd.to_numeric(feature_store[col], downcast="integer")
+```
+
+**Bénéfice** : RAM 600-900 MB au lieu de 1-1.8 GB, `.loc` plus rapide.
+
+**Risque** : faible si LightGBM a été entraîné sur float32 (à vérifier ;
+si entraîné float64, faible perte de précision sur les agrégats — à
+mesurer via test de parité).
+
+#### E. Précompute numpy template → latence -5 à -15 ms par requête
+
+Le hot path actuel fait `pd.DataFrame([raw])` + `pd.Categorical` × 14 +
+`get_dummies` + `concat` + `reindex(768)` + `replace(inf, nan)` pour
+chaque requête. Plan : remplacer par une copie de template numpy pré-aligné
+sur `feature_names`, avec un mapping `name → index` calculé une fois au boot.
+
+**Bénéfice** : élimine pandas du hot path, latence p50 probablement de
+20-40 ms à 3-8 ms.
+
+**Risque** : moyen. Demande un golden-file test exhaustif vs assemble()
+actuel (sortie strictement identique sur ≥100 cas couvrant les deux
+branches connu/inconnu, toutes les catégorielles).
+
+#### F. Sécurité advisory (non bloquant pour formation OC)
+
+- **Rate limiting** : `slowapi` per-IP sur `/predict` (10 req/min).
+- **CORS** : `CORSMiddleware` avec `allow_origins=["https://...your-frontend..."]`.
+- **API key** : header `X-API-Key` via dépendance FastAPI si besoin
+  d'auth simple.
+- **Non-root user dans Dockerfile** : `RUN useradd -m appuser && USER appuser`
+  (HF Spaces enforce déjà un user namespace, mais defense-in-depth).
+
+À faire **uniquement** si l'API doit être exposée à du trafic réel.
+
+### Findings non-bloquants laissés tels quels
+
+- **Chemins absolus** dans `scripts/build_feature_store.py:41`,
+  `scripts/export_model.py:27-30`, `scripts/check_registry.py:4` —
+  scripts personnels one-shot, pas de gain immédiat à env-var-iser.
+- **CI action versions** (`actions/checkout@v6`, `setup-python@v6`,
+  `upload-artifact@v7`) — flagged par code-reviewer comme non-existantes,
+  mais le CI tourne, donc soit elles existent, soit GitHub résout
+  silencieusement. Non-bloquant.
+- **Duplication ratios** offline / runtime (orchestrator vs `api/ratios.py`)
+  — intentionnelle et confirmée par 2 reviewers (le runtime a besoin du
+  scrub `inf → NaN` que le offline n'a pas).
+- **`joblib.load` pickle-based** — modèle baked-in à l'image build,
+  pas un vecteur d'attaque exploitable.

.claude/settings.local.json

@@ -5,7 +5,13 @@
       "Bash(python -c ' *)",
       "Bash(uv run --active python -c ' *)",
       "WebFetch(domain:github.com)",
-      "WebSearch"
+      "WebSearch",
+      "Bash(uv run *)",
+      "Bash(git -C \"C:\\\\Users\\\\Kevin\\\\projects\\\\OC_P8\" diff --staged)",
+      "Bash(git -C \"C:\\\\Users\\\\Kevin\\\\projects\\\\OC_P8\" diff)",
+      "Bash(git -C \"C:\\\\Users\\\\Kevin\\\\projects\\\\OC_P8\" log --oneline -10)",
+      "PowerShell(uv run ruff check api feature_engineering tests scripts)",
+      "PowerShell(uv run pytest --cov=api --cov-fail-under=80 -q)"
     ]
   }
 }

README.md

@@ -18,7 +18,7 @@ pinned: false
 <div align="center">
   <h1 align="center">OC P8 — Credit Scoring API</h1>
   <p align="center">
-    Production-grade FastAPI wrapper around the XGBoost credit scoring model
+    Production-grade FastAPI wrapper around the LightGBM credit scoring model
     trained in OC_P6. Built for <em>Prêt à Dépenser</em>'s Crédit Express department:
     real-time default risk prediction for loan officers.
     <br />
@@ -63,7 +63,7 @@ pinned: false
 The **Credit Scoring API** exposes a single `POST /predict` endpoint. Given a loan application (`SK_ID_CURR` + 120 raw `application_train` fields), it returns:
 
 - `probability_default` — model score between 0 and 1
-- `decision` — `true` (loan refused) if `proba ≥ 0.33`, `false` (loan granted) otherwise
+- `decision` — `"REFUSED"` if `proba ≥ 0.33`, `"GRANTED"` otherwise
 - `threshold`, `model_version`, `client_known` — explainability metadata
 
 The threshold **0.33** is optimised for an asymmetric cost function (10 × false negatives + false positives), meaning the model is intentionally conservative: missing a bad borrower costs 10× more than wrongly refusing a good one.
@@ -74,7 +74,7 @@ The threshold **0.33** is optimised for an asymmetric cost function (10 × false
 
 [![Python][python-badge]][python-url]
 [![FastAPI][fastapi-badge]][fastapi-url]
-[![XGBoost][xgboost-badge]][xgboost-url]
+[![LightGBM][lightgbm-badge]][lightgbm-url]
 [![uv][uv-badge]][uv-url]
 [![Docker][docker-badge]][docker-url]
 [![GitHub Actions][gha-badge]][gha-url]
@@ -199,14 +199,14 @@ Example response:
 {
   "sk_id_curr": 100001,
   "probability_default": 0.1523,
-  "decision": false,
+  "decision": "GRANTED",
   "threshold": 0.33,
-  "model_version": "xgb-v1.0",
+  "model_version": "2",
   "client_known": true
 }
 ```
 
-`decision: false` = loan **granted** · `decision: true` = loan **refused**
+`decision: "GRANTED"` = loan **granted** · `decision: "REFUSED"` = loan **refused**
 
 You can also use the interactive **Swagger UI** at `/docs` → `POST /predict` → **Try it out**.
 
@@ -249,7 +249,7 @@ JSON {SK_ID_CURR + 120 raw application_train fields}
 | **Known client** | `SK_ID_CURR` found in `features_store.parquet` | Pre-computed bureau / prev / POS / CC / install |
 | **Unknown client** | `SK_ID_CURR` not found | `no_history_template.json` (counts=0, rest NaN) |
 
-The unknown-client path preserves XGBoost's training-time NaN signal ("no historical data") rather than imputing fictitious medians.
+The unknown-client path preserves LightGBM's training-time NaN signal ("no historical data") rather than imputing fictitious medians.
 
 ### Data layer — code/data separation
 
@@ -408,8 +408,8 @@ Internal project — Prêt à Dépenser MLOps formation OpenClassrooms.
 [python-url]: https://www.python.org/
 [fastapi-badge]: https://img.shields.io/badge/FastAPI-0.115-009688?style=for-the-badge&logo=fastapi&logoColor=white
 [fastapi-url]: https://fastapi.tiangolo.com/
-[xgboost-badge]: https://img.shields.io/badge/XGBoost-2.x-F7931E?style=for-the-badge
-[xgboost-url]: https://xgboost.readthedocs.io/
+[lightgbm-badge]: https://img.shields.io/badge/LightGBM-4.x-2E8B57?style=for-the-badge
+[lightgbm-url]: https://lightgbm.readthedocs.io/
 [uv-badge]: https://img.shields.io/badge/uv-package%20manager-DE5FE9?style=for-the-badge
 [uv-url]: https://docs.astral.sh/uv/
 [docker-badge]: https://img.shields.io/badge/Docker-container-2496ED?style=for-the-badge&logo=docker&logoColor=white

api/inference_assembler.py

@@ -90,8 +90,7 @@ def assemble(
 
     # 2. Aggregate portion — lookup or template.
     if sk_id_curr in artefacts.feature_store.index:
-        agg_part = artefacts.feature_store.loc[[sk_id_curr]].copy()
-        agg_part.reset_index(drop=True, inplace=True)
+        agg_part = artefacts.feature_store.loc[[sk_id_curr]].reset_index(drop=True)
         client_known = True
     else:
         agg_part = pd.DataFrame([artefacts.no_history_template])

api/main.py

@@ -112,9 +112,8 @@ async def unhandled_exception_handler(request: Request, exc: Exception) -> JSONR
     )
 
 
-
-@app.get("/") # La page d'accueil 
-async def read_root():
+@app.get("/", tags=["meta"])
+async def read_root() -> dict[str, str]:
     return {"message": "Welcome to the CREDIT DEFAULT predictor API for Prêt à Dépenser"}
 
 

api/predictor.py

@@ -10,13 +10,12 @@ from __future__ import annotations
 
 import json
 from pathlib import Path
-from typing import Literal
 
 import joblib
 import numpy as np
 import pandas as pd
 
-Decision = Literal["GRANTED", "REFUSED"]
+from api.schemas import Decision
 
 
 class CreditScoringPredictor:

api/schemas.py

@@ -12,7 +12,7 @@ nonsensical inputs at the API boundary.
 
 from __future__ import annotations
 
-from typing import Literal, Optional
+from typing import Literal
 
 from pydantic import BaseModel, ConfigDict, Field
 
@@ -178,9 +178,9 @@ class PredictionRequest(BaseModel):
     CNT_CHILDREN: int = Field(ge=0, le=20)
     AMT_INCOME_TOTAL: float = Field(gt=0)
     AMT_CREDIT: float = Field(gt=0)
-    AMT_ANNUITY: Optional[float] = Field(default=None, gt=0)
-    AMT_GOODS_PRICE: Optional[float] = Field(default=None, gt=0)
-    NAME_TYPE_SUITE: Optional[TypeSuite] = None
+    AMT_ANNUITY: float | None = Field(default=None, gt=0)
+    AMT_GOODS_PRICE: float | None = Field(default=None, gt=0)
+    NAME_TYPE_SUITE: TypeSuite | None = None
     NAME_INCOME_TYPE: IncomeType
     NAME_EDUCATION_TYPE: EducationType
     NAME_FAMILY_STATUS: FamilyStatus
@@ -194,14 +194,14 @@ class PredictionRequest(BaseModel):
     )
     DAYS_REGISTRATION: float = Field(le=0, ge=-25000)
     DAYS_ID_PUBLISH: int = Field(le=0, ge=-10000)
-    OWN_CAR_AGE: Optional[float] = Field(default=None, ge=0, le=100)
+    OWN_CAR_AGE: float | None = Field(default=None, ge=0, le=100)
     FLAG_MOBIL: int = Field(ge=0, le=1)
     FLAG_EMP_PHONE: int = Field(ge=0, le=1)
     FLAG_WORK_PHONE: int = Field(ge=0, le=1)
     FLAG_CONT_MOBILE: int = Field(ge=0, le=1)
     FLAG_PHONE: int = Field(ge=0, le=1)
     FLAG_EMAIL: int = Field(ge=0, le=1)
-    OCCUPATION_TYPE: Optional[OccupationType] = None
+    OCCUPATION_TYPE: OccupationType | None = None
     CNT_FAM_MEMBERS: float = Field(ge=1, le=20)
     REGION_RATING_CLIENT: int = Field(ge=1, le=3)
     REGION_RATING_CLIENT_W_CITY: int = Field(ge=1, le=3)
@@ -216,64 +216,64 @@ class PredictionRequest(BaseModel):
     ORGANIZATION_TYPE: OrganizationType
 
     # External scoring sources ----------------------------------------------
-    EXT_SOURCE_1: Optional[float] = Field(default=None, ge=0, le=1)
-    EXT_SOURCE_2: Optional[float] = Field(default=None, ge=0, le=1)
-    EXT_SOURCE_3: Optional[float] = Field(default=None, ge=0, le=1)
+    EXT_SOURCE_1: float | None = Field(default=None, ge=0, le=1)
+    EXT_SOURCE_2: float | None = Field(default=None, ge=0, le=1)
+    EXT_SOURCE_3: float | None = Field(default=None, ge=0, le=1)
 
     # Building characteristics (mostly nullable, ratios in [0, 1]) ----------
-    APARTMENTS_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    BASEMENTAREA_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    YEARS_BEGINEXPLUATATION_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    YEARS_BUILD_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    COMMONAREA_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    ELEVATORS_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    ENTRANCES_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    FLOORSMAX_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    FLOORSMIN_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    LANDAREA_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    LIVINGAPARTMENTS_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    LIVINGAREA_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    NONLIVINGAPARTMENTS_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    NONLIVINGAREA_AVG: Optional[float] = Field(default=None, ge=0, le=1)
-    APARTMENTS_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    BASEMENTAREA_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    YEARS_BEGINEXPLUATATION_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    YEARS_BUILD_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    COMMONAREA_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    ELEVATORS_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    ENTRANCES_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    FLOORSMAX_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    FLOORSMIN_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    LANDAREA_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    LIVINGAPARTMENTS_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    LIVINGAREA_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    NONLIVINGAPARTMENTS_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    NONLIVINGAREA_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    APARTMENTS_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    BASEMENTAREA_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    YEARS_BEGINEXPLUATATION_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    YEARS_BUILD_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    COMMONAREA_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    ELEVATORS_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    ENTRANCES_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    FLOORSMAX_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    FLOORSMIN_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    LANDAREA_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    LIVINGAPARTMENTS_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    LIVINGAREA_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    NONLIVINGAPARTMENTS_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    NONLIVINGAREA_MEDI: Optional[float] = Field(default=None, ge=0, le=1)
-    FONDKAPREMONT_MODE: Optional[FondKapremontMode] = None
-    HOUSETYPE_MODE: Optional[HouseTypeMode] = None
-    TOTALAREA_MODE: Optional[float] = Field(default=None, ge=0, le=1)
-    WALLSMATERIAL_MODE: Optional[WallsMaterialMode] = None
-    EMERGENCYSTATE_MODE: Optional[EmergencyStateMode] = None
+    APARTMENTS_AVG: float | None = Field(default=None, ge=0, le=1)
+    BASEMENTAREA_AVG: float | None = Field(default=None, ge=0, le=1)
+    YEARS_BEGINEXPLUATATION_AVG: float | None = Field(default=None, ge=0, le=1)
+    YEARS_BUILD_AVG: float | None = Field(default=None, ge=0, le=1)
+    COMMONAREA_AVG: float | None = Field(default=None, ge=0, le=1)
+    ELEVATORS_AVG: float | None = Field(default=None, ge=0, le=1)
+    ENTRANCES_AVG: float | None = Field(default=None, ge=0, le=1)
+    FLOORSMAX_AVG: float | None = Field(default=None, ge=0, le=1)
+    FLOORSMIN_AVG: float | None = Field(default=None, ge=0, le=1)
+    LANDAREA_AVG: float | None = Field(default=None, ge=0, le=1)
+    LIVINGAPARTMENTS_AVG: float | None = Field(default=None, ge=0, le=1)
+    LIVINGAREA_AVG: float | None = Field(default=None, ge=0, le=1)
+    NONLIVINGAPARTMENTS_AVG: float | None = Field(default=None, ge=0, le=1)
+    NONLIVINGAREA_AVG: float | None = Field(default=None, ge=0, le=1)
+    APARTMENTS_MODE: float | None = Field(default=None, ge=0, le=1)
+    BASEMENTAREA_MODE: float | None = Field(default=None, ge=0, le=1)
+    YEARS_BEGINEXPLUATATION_MODE: float | None = Field(default=None, ge=0, le=1)
+    YEARS_BUILD_MODE: float | None = Field(default=None, ge=0, le=1)
+    COMMONAREA_MODE: float | None = Field(default=None, ge=0, le=1)
+    ELEVATORS_MODE: float | None = Field(default=None, ge=0, le=1)
+    ENTRANCES_MODE: float | None = Field(default=None, ge=0, le=1)
+    FLOORSMAX_MODE: float | None = Field(default=None, ge=0, le=1)
+    FLOORSMIN_MODE: float | None = Field(default=None, ge=0, le=1)
+    LANDAREA_MODE: float | None = Field(default=None, ge=0, le=1)
+    LIVINGAPARTMENTS_MODE: float | None = Field(default=None, ge=0, le=1)
+    LIVINGAREA_MODE: float | None = Field(default=None, ge=0, le=1)
+    NONLIVINGAPARTMENTS_MODE: float | None = Field(default=None, ge=0, le=1)
+    NONLIVINGAREA_MODE: float | None = Field(default=None, ge=0, le=1)
+    APARTMENTS_MEDI: float | None = Field(default=None, ge=0, le=1)
+    BASEMENTAREA_MEDI: float | None = Field(default=None, ge=0, le=1)
+    YEARS_BEGINEXPLUATATION_MEDI: float | None = Field(default=None, ge=0, le=1)
+    YEARS_BUILD_MEDI: float | None = Field(default=None, ge=0, le=1)
+    COMMONAREA_MEDI: float | None = Field(default=None, ge=0, le=1)
+    ELEVATORS_MEDI: float | None = Field(default=None, ge=0, le=1)
+    ENTRANCES_MEDI: float | None = Field(default=None, ge=0, le=1)
+    FLOORSMAX_MEDI: float | None = Field(default=None, ge=0, le=1)
+    FLOORSMIN_MEDI: float | None = Field(default=None, ge=0, le=1)
+    LANDAREA_MEDI: float | None = Field(default=None, ge=0, le=1)
+    LIVINGAPARTMENTS_MEDI: float | None = Field(default=None, ge=0, le=1)
+    LIVINGAREA_MEDI: float | None = Field(default=None, ge=0, le=1)
+    NONLIVINGAPARTMENTS_MEDI: float | None = Field(default=None, ge=0, le=1)
+    NONLIVINGAREA_MEDI: float | None = Field(default=None, ge=0, le=1)
+    FONDKAPREMONT_MODE: FondKapremontMode | None = None
+    HOUSETYPE_MODE: HouseTypeMode | None = None
+    TOTALAREA_MODE: float | None = Field(default=None, ge=0, le=1)
+    WALLSMATERIAL_MODE: WallsMaterialMode | None = None
+    EMERGENCYSTATE_MODE: EmergencyStateMode | None = None
 
     # Social circle ---------------------------------------------------------
-    OBS_30_CNT_SOCIAL_CIRCLE: Optional[float] = Field(default=None, ge=0, le=500)
-    DEF_30_CNT_SOCIAL_CIRCLE: Optional[float] = Field(default=None, ge=0, le=500)
-    OBS_60_CNT_SOCIAL_CIRCLE: Optional[float] = Field(default=None, ge=0, le=500)
-    DEF_60_CNT_SOCIAL_CIRCLE: Optional[float] = Field(default=None, ge=0, le=500)
+    OBS_30_CNT_SOCIAL_CIRCLE: float | None = Field(default=None, ge=0, le=500)
+    DEF_30_CNT_SOCIAL_CIRCLE: float | None = Field(default=None, ge=0, le=500)
+    OBS_60_CNT_SOCIAL_CIRCLE: float | None = Field(default=None, ge=0, le=500)
+    DEF_60_CNT_SOCIAL_CIRCLE: float | None = Field(default=None, ge=0, le=500)
 
     DAYS_LAST_PHONE_CHANGE: float = Field(le=0, ge=-15000)
 
@@ -300,12 +300,12 @@ class PredictionRequest(BaseModel):
     FLAG_DOCUMENT_21: int = Field(ge=0, le=1)
 
     # Credit bureau request volume -----------------------------------------
-    AMT_REQ_CREDIT_BUREAU_HOUR: Optional[float] = Field(default=None, ge=0, le=500)
-    AMT_REQ_CREDIT_BUREAU_DAY: Optional[float] = Field(default=None, ge=0, le=500)
-    AMT_REQ_CREDIT_BUREAU_WEEK: Optional[float] = Field(default=None, ge=0, le=500)
-    AMT_REQ_CREDIT_BUREAU_MON: Optional[float] = Field(default=None, ge=0, le=500)
-    AMT_REQ_CREDIT_BUREAU_QRT: Optional[float] = Field(default=None, ge=0, le=500)
-    AMT_REQ_CREDIT_BUREAU_YEAR: Optional[float] = Field(default=None, ge=0, le=500)
+    AMT_REQ_CREDIT_BUREAU_HOUR: float | None = Field(default=None, ge=0, le=500)
+    AMT_REQ_CREDIT_BUREAU_DAY: float | None = Field(default=None, ge=0, le=500)
+    AMT_REQ_CREDIT_BUREAU_WEEK: float | None = Field(default=None, ge=0, le=500)
+    AMT_REQ_CREDIT_BUREAU_MON: float | None = Field(default=None, ge=0, le=500)
+    AMT_REQ_CREDIT_BUREAU_QRT: float | None = Field(default=None, ge=0, le=500)
+    AMT_REQ_CREDIT_BUREAU_YEAR: float | None = Field(default=None, ge=0, le=500)
 
 
 # ---------------------------------------------------------------------------

api/settings.py

@@ -40,4 +40,6 @@ HF_DATASET_REPO_ID = os.getenv("OC_P8_HF_DATASET_REPO_ID", "KLEB38/oc-p8-feature
 HF_DATASET_FILENAME = os.getenv("OC_P8_HF_DATASET_FILENAME", "features_store.parquet")
 
 # Default fallback if model_info.json does not expose the optimised threshold.
+# 0.33 minimises the business cost function 10*FN + FP from OC_P6 — re-run the
+# threshold search if the model is retrained.
 DEFAULT_THRESHOLD = 0.33

feature_engineering/aggregations.py

@@ -198,8 +198,8 @@ def installments_payments(
 
     ins["DPD"] = ins["DAYS_ENTRY_PAYMENT"] - ins["DAYS_INSTALMENT"]
     ins["DBD"] = ins["DAYS_INSTALMENT"] - ins["DAYS_ENTRY_PAYMENT"]
-    ins["DPD"] = ins["DPD"].apply(lambda x: x if x > 0 else 0)
-    ins["DBD"] = ins["DBD"].apply(lambda x: x if x > 0 else 0)
+    ins["DPD"] = ins["DPD"].clip(lower=0)
+    ins["DBD"] = ins["DBD"].clip(lower=0)
 
     aggregations: dict[str, list[str]] = {
         "NUM_INSTALMENT_VERSION": ["nunique"],

feature_engineering/orchestrator.py

@@ -8,6 +8,7 @@ the model.
 from __future__ import annotations
 
 import gc
+import logging
 from pathlib import Path
 
 import numpy as np
@@ -22,6 +23,8 @@ from feature_engineering.aggregations import (
     previous_applications,
 )
 
+logger = logging.getLogger(__name__)
+
 
 def app_train_clean(
     data_dir: Path, num_rows: int | None = None, nan_as_category: bool = False
@@ -56,31 +59,31 @@ def merge_files(data_dir: Path, debug: bool = False) -> pd.DataFrame:
     df = app_train_clean(data_dir, num_rows)
 
     bureau = bureau_and_balance(data_dir, num_rows)
-    print("Bureau df shape:", bureau.shape)
+    logger.info("Bureau df shape: %s", bureau.shape)
     df = df.join(bureau, how="left", on="SK_ID_CURR")
     del bureau
     gc.collect()
 
     prev = previous_applications(data_dir, num_rows)
-    print("Previous applications df shape:", prev.shape)
+    logger.info("Previous applications df shape: %s", prev.shape)
     df = df.join(prev, how="left", on="SK_ID_CURR")
     del prev
     gc.collect()
 
     pos = pos_cash(data_dir, num_rows)
-    print("Pos-cash balance df shape:", pos.shape)
+    logger.info("Pos-cash balance df shape: %s", pos.shape)
     df = df.join(pos, how="left", on="SK_ID_CURR")
     del pos
     gc.collect()
 
     ins = installments_payments(data_dir, num_rows)
-    print("Installments payments df shape:", ins.shape)
+    logger.info("Installments payments df shape: %s", ins.shape)
     df = df.join(ins, how="left", on="SK_ID_CURR")
     del ins
     gc.collect()
 
     cc = credit_card_balance(data_dir, num_rows)
-    print("Credit card balance df shape:", cc.shape)
+    logger.info("Credit card balance df shape: %s", cc.shape)
     df = df.join(cc, how="left", on="SK_ID_CURR")
     del cc
     gc.collect()

scripts/build_no_history_template.py

@@ -35,10 +35,10 @@ def main() -> None:
     feature_names = json.loads(FEATURE_NAMES_PATH.read_text())
     store = pd.read_parquet(FEATURE_STORE_PATH)
 
-    aggregate_cols = [c for c in store.columns]
+    aggregate_cols = list(store.columns)
     print(f"Inspecting {len(aggregate_cols)} aggregate columns...")
 
-    template: dict[str, float | int] = {}
+    template: dict[str, float | int | None] = {}
     count_cols: list[str] = []
     nan_cols: list[str] = []
 
@@ -56,12 +56,12 @@ def main() -> None:
     print("Sample NaN columns:  ", nan_cols[:5])
 
     OUT_TEMPLATE.write_text(json.dumps(template, indent=2))
-    print(f"\n✅ {OUT_TEMPLATE} ({len(template)} entries)")
+    print(f"\n[OK] {OUT_TEMPLATE} ({len(template)} entries)")
 
     # Sanity check: every aggregate column from the store is also in feature_names
     missing = [c for c in aggregate_cols if c not in feature_names]
     if missing:
-        print(f"⚠️  {len(missing)} columns in parquet but absent from feature_names")
+        print(f"[WARN] {len(missing)} columns in parquet but absent from feature_names")
         print(f"   first 5: {missing[:5]}")
 
 

scripts/export_model.py

@@ -52,6 +52,9 @@ print(f"  Version   : {mv.version}")
 print(f"  Run ID    : {mv.run_id}")
 print(f"  Source    : {mv.source}")
 
+if mv.run_id is None:
+    raise RuntimeError(f"Model version {MODEL_NAME}/{MODEL_VERSION} has no associated run_id")
+
 run = client.get_run(mv.run_id)
 print("\n  Run metrics:")
 for k, v in run.data.metrics.items():
@@ -72,7 +75,7 @@ model = mlflow.pyfunc.load_model(str(P6_MODEL_PATH))
 print("\n  Model signature:")
 sig = model.metadata.signature
 if sig is None:
-    print("    ⚠️  No signature logged - inputs will not be validated at inference!")
+    print("    [WARN] No signature logged - inputs will not be validated at inference!")
     n_features = None
     feature_names = None
 else:
@@ -87,7 +90,7 @@ else:
 # ============================================================
 joblib.dump(model, OUTPUT_MODEL)
 size_mb = OUTPUT_MODEL.stat().st_size / 1e6
-print(f"\n✅ Model saved to: {OUTPUT_MODEL} ({size_mb:.2f} MB)")
+print(f"\n[OK] Model saved to: {OUTPUT_MODEL} ({size_mb:.2f} MB)")
 
 # ============================================================
 # 4. Save provenance metadata
@@ -106,6 +109,6 @@ info = {
 }
 with open(OUTPUT_INFO, "w") as f:
     json.dump(info, f, indent=2, default=str)
-print(f"✅ Metadata saved to: {OUTPUT_INFO}")
+print(f"[OK] Metadata saved to: {OUTPUT_INFO}")
 
 print("=" * 60)