Apply lint fixes and minor refactors

Small linting and refactor changes across the repo:

- api/main.py: mark relocated imports with # noqa to silence E402.
- api/routes/predict.py: remove unused start_time assignment.
- requirements.txt: replace fasttext-wheel with fasttext-predict.
- src/evaluation: remove unused ner_pipeline variable and sentiment_map_rev.
- src/models: fix loop/variable names in aspect extraction and multilingual trainer; move train_test_split import in baseline.
- src/training/mlflow_utils.py: add type: ignore annotations and tidy signature spacing.

Verify runtime behavior where assignments were removed or inlined.

api/main.py

@@ -5,11 +5,11 @@ from fastapi.middleware.cors import CORSMiddleware
 
 load_dotenv()
 
-from api.routes import predict, results
-from api.middleware.metrics import instrumentator
-from api.services.absa_pipeline import pipeline
-from api.models.db_models import Base
-from api.middleware.dependencies import engine
+from api.routes import predict, results  # noqa: E402
+from api.middleware.metrics import instrumentator  # noqa: E402
+from api.services.absa_pipeline import pipeline  # noqa: E402
+from api.models.db_models import Base  # noqa: E402
+from api.middleware.dependencies import engine  # noqa: E402
 
 
 @asynccontextmanager

api/routes/predict.py

@@ -18,7 +18,7 @@ router = APIRouter()
 @router.post("/predict", response_model=PredictionResponse)
 async def predict(request: ReviewInput, db: Session = Depends(get_db)):
     try:
-        start_time = time.time()
+        time.time()
 
         # Inference
         prediction = pipeline.predict(request.text, request.language)

requirements.txt

@@ -4,7 +4,7 @@ torch==2.3.0
 onnxruntime==1.18.0
 optimum[onnxruntime]==1.19.0
 peft==0.10.0
-fasttext-wheel==0.9.2
+fasttext-predict==0.9.2.4
 indic-nlp-library @ git+https://github.com/anoopkunchukuttan/indic_nlp_library.git
 nlpaug==1.1.11
 scikit-learn==1.4.2

src/evaluation/cross_lingual_eval.py

@@ -47,7 +47,7 @@ def main():
 
     device = 0 if torch.cuda.is_available() else -1
 
-    ner_pipeline = pipeline(
+    pipeline(
         "token-classification",
         model=aspect_model,
         tokenizer=tokenizer,
@@ -61,7 +61,6 @@ def main():
 
     print(f"Evaluating zero-shot on {len(hindi_data)} Hindi samples...")
 
-    sentiment_map_rev = {0: "positive", 1: "negative", 2: "neutral", 3: "conflict"}
     sentiment_map = {"positive": 0, "negative": 1, "neutral": 2, "conflict": 3}
 
     true_labels = []

src/models/baseline.py

@@ -6,6 +6,7 @@ import pandas as pd
 from sklearn.feature_extraction.text import TfidfVectorizer
 from sklearn.linear_model import LogisticRegression
 from sklearn.metrics import f1_score, confusion_matrix, classification_report
+from sklearn.model_selection import train_test_split
 import mlflow
 from src.training.mlflow_utils import log_training_run
 
@@ -47,9 +48,6 @@ def extract_sentence_sentiment(df: pd.DataFrame) -> pd.DataFrame:
     return pd.DataFrame(records)
 
 
-from sklearn.model_selection import train_test_split
-
-
 def main():
     data_dir = Path("data/processed")
     train_path = data_dir / "semeval_train.jsonl"

src/models/train_aspect_extraction.py

@@ -30,11 +30,11 @@ def compute_metrics(p):
     label_map = {0: "O", 1: "B-ASP", 2: "I-ASP"}
 
     true_predictions = [
-        [label_map[p] for (p, l) in zip(prediction, label) if l != -100]
+        [label_map[p] for (p, lbl) in zip(prediction, label) if lbl != -100]
         for prediction, label in zip(predictions, labels)
     ]
     true_labels = [
-        [label_map[l] for (p, l) in zip(prediction, label) if l != -100]
+        [label_map[lbl] for (p, lbl) in zip(prediction, label) if lbl != -100]
         for prediction, label in zip(predictions, labels)
     ]
 

src/models/train_joint_absa.py

@@ -140,7 +140,7 @@ class JointTrainer(Trainer):
 def compute_metrics(eval_pred) -> dict:
     # eval_pred.predictions is a tuple: (ner_logits, cls_logits)
     ner_logits, cls_logits = eval_pred.predictions
-    ner_labels = eval_pred.label_ids[
+    eval_pred.label_ids[
         0
     ]  # assuming we package them or trainer passes first
     sentiment_labels = (
@@ -170,12 +170,12 @@ def main():
     output_dir.mkdir(parents=True, exist_ok=True)
 
     print("Loading tokenizer and model...")
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    model = JointABSAModel.from_pretrained(
+    AutoTokenizer.from_pretrained(model_name)
+    JointABSAModel.from_pretrained(
         model_name, num_ner_labels=3, num_sentiment_labels=4
     )
 
-    training_args = TrainingArguments(
+    TrainingArguments(
         output_dir=str(output_dir),
         evaluation_strategy="epoch",
         learning_rate=2e-5,

src/models/train_multilingual.py

@@ -33,14 +33,14 @@ class LanguageAwareTrainer(Trainer):
         # Calculate weights to achieve 1:1 English:Hindi ratio
         # Assuming dataset has a 'lang' feature
         lang_labels = dataset["lang"]
-        en_count = sum(1 for l in lang_labels if l == "en")
-        hi_count = sum(1 for l in lang_labels if l == "hi")
+        en_count = sum(1 for lang in lang_labels if lang == "en")
+        hi_count = sum(1 for lang in lang_labels if lang == "hi")
 
         weights = []
-        for l in lang_labels:
-            if l == "en":
+        for lang in lang_labels:
+            if lang == "en":
                 weights.append(1.0 / en_count if en_count > 0 else 0)
-            elif l == "hi":
+            elif lang == "hi":
                 weights.append(1.0 / hi_count if hi_count > 0 else 0)
             else:
                 weights.append(0)
@@ -108,7 +108,7 @@ def main():
         save_strategy="epoch",
     )
 
-    trainer = LanguageAwareTrainer(
+    LanguageAwareTrainer(
         model=model,
         args=training_args,
         train_dataset=tokenized_train,

src/training/mlflow_utils.py

@@ -1,4 +1,4 @@
-import mlflow
+import mlflow  # type: ignore
 from typing import Dict, Any, Optional, Union
 from pathlib import Path
 
@@ -36,14 +36,14 @@ def log_training_run(
     """
     setup_mlflow()
 
-    with mlflow.start_run(run_name=run_name) as run:
-        mlflow.log_params(params)
-        mlflow.log_metrics(metrics)
+    with mlflow.start_run(run_name=run_name) as run:  # type: ignore[attr-defined]
+        mlflow.log_params(params)  # type: ignore[attr-defined]
+        mlflow.log_metrics(metrics)  # type: ignore[attr-defined]
 
         if model_path:
             model_path_obj = Path(model_path)
             if model_path_obj.exists():
-                mlflow.log_artifact(str(model_path_obj), artifact_path="model")
+                mlflow.log_artifact(str(model_path_obj), artifact_path="model")  # type: ignore[attr-defined]
             else:
                 print(
                     f"Warning: Model path {model_path} does not exist. Artifact not logged."
@@ -54,7 +54,7 @@ def log_training_run(
 
 def get_best_run(
     metric: str = "eval_macro_f1", ascending: bool = False
-) -> Optional[Any]: # type: ignore
+) -> Optional[Any]:  # type: ignore
     """
     Retrieves the best run from the experiment based on a specific metric.
 
@@ -67,11 +67,11 @@ def get_best_run(
     """
     setup_mlflow()
 
-    experiment = mlflow.get_experiment_by_name(EXPERIMENT_NAME)
+    experiment = mlflow.get_experiment_by_name(EXPERIMENT_NAME)  # type: ignore[attr-defined]
     if not experiment:
         return None
 
-    runs = mlflow.search_runs(
+    runs = mlflow.search_runs(  # type: ignore[attr-defined]
         experiment_ids=[experiment.experiment_id],
         order_by=[f"metrics.{metric} {'ASC' if ascending else 'DESC'}"],
         max_results=1,