diff --git a/src/osc_transformer_based_extractor/relevance_detector/fine_tune.py b/src/osc_transformer_based_extractor/relevance_detector/fine_tune.py
index 5b85193..456e0d8 100644
--- a/src/osc_transformer_based_extractor/relevance_detector/fine_tune.py
+++ b/src/osc_transformer_based_extractor/relevance_detector/fine_tune.py
@@ -13,6 +13,8 @@
 import shutil
 import pandas as pd
 import torch
+import json
+import time
 from transformers import (
     TrainingArguments,
     Trainer,
@@ -20,7 +22,7 @@
     AutoTokenizer,
 )
 from sklearn.model_selection import train_test_split
-from sklearn.metrics import accuracy_score
+from sklearn.metrics import accuracy_score, precision_recall_fscore_support
 from torch.utils.data import Dataset
 
 
@@ -219,8 +221,23 @@ def fine_tune_model(
     checkpoint_dir = os.path.join(saved_model_path, "checkpoints")
     os.makedirs(saved_model_path, exist_ok=True)
 
+    def compute_metrics(pred):
+        predictions, labels = pred
+        predictions = predictions.argmax(axis=-1)
+        precision, recall, f1, _ = precision_recall_fscore_support(
+            labels, predictions, average="weighted"
+        )
+        acc = accuracy_score(labels, predictions)
+        return {
+            "accuracy": acc,
+            "f1": f1,
+            "precision": precision,
+            "recall": recall,
+        }
+
     training_args = TrainingArguments(
         output_dir=checkpoint_dir,
+        dataloader_pin_memory=False,
         evaluation_strategy="epoch",  # Evaluate at the end of each epoch
         logging_dir="./logs",  # Directory for logs
         logging_steps=10,  # Log every 10 steps
@@ -236,6 +253,7 @@ def fine_tune_model(
         metric_for_best_model="eval_loss",
         greater_is_better=False,
         save_total_limit=1,
+        seed=42,
     )
 
     # Initialize Trainer
@@ -244,11 +262,14 @@ def fine_tune_model(
         args=training_args,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
+        compute_metrics=compute_metrics,
     )
 
+    start = time.time()
     # Start Training
     trainer.train()
-
+    end = time.time()
+    print("Time needed to train model =", end - start)
     # Save the final trained model and config
     trainer.save_model(saved_model_path)
 
@@ -261,6 +282,12 @@ def fine_tune_model(
     for key, value in eval_result.items():
         print(f"{key}: {value}")
 
+    with open(
+        os.path.join(output_dir, f"evaluation_results_{export_model_name}.json"), "w"
+    ) as json_file:
+        json.dump(eval_result, json_file, indent=4)
+    print("Evaluation results saved to 'evaluation_results.json'")
+
     # Predict labels for the evaluation dataset
     predictions = trainer.predict(eval_dataset)
     predicted_labels = predictions.predictions.argmax(axis=1)
diff --git a/src/osc_transformer_based_extractor/relevance_detector/inference.py b/src/osc_transformer_based_extractor/relevance_detector/inference.py
index 6e007dc..095880f 100644
--- a/src/osc_transformer_based_extractor/relevance_detector/inference.py
+++ b/src/osc_transformer_based_extractor/relevance_detector/inference.py
@@ -7,6 +7,43 @@
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, AutoConfig
 
 
+def combine_and_filter_xlsx_files(folder_path, output_file):
+    """
+    Combine all .xlsx files in a folder, filter rows where paragraph_relevance_flag is 1,
+    and save the result as a new .xlsx file.
+
+    Parameters:
+    folder_path (str): Path to the folder containing .xlsx files.
+    output_file (str): Path to save the filtered combined file.
+    """
+    all_dataframes = []
+
+    # Iterate through all files in the folder
+    for file_name in os.listdir(folder_path):
+        if file_name.endswith(".xlsx"):
+            file_path = os.path.join(folder_path, file_name)
+            try:
+                # Read the Excel file into a DataFrame
+                df = pd.read_excel(file_path)
+                all_dataframes.append(df)
+            except Exception as e:
+                print(f"Error reading {file_name}: {e}")
+
+    # Combine all DataFrames into one
+    if all_dataframes:
+        combined_df = pd.concat(all_dataframes, axis=0, ignore_index=True)
+
+        # Filter rows where paragraph_relevance_flag is 1
+        filtered_df = combined_df[combined_df["paragraph_relevance_flag"] == 1]
+
+        # Save the filtered DataFrame to an Excel file
+        file_name = "combined_inference.xlsx"
+        filtered_df.to_excel(os.path.join(output_file, file_name), index=False)
+        print(f"Filtered data saved to {output_file}")
+    else:
+        print("No valid .xlsx files found in the folder.")
+
+
 def validate_path_exists(path: str, which_path: str):
     """
     Validate if the given path exists.
@@ -170,3 +207,11 @@ def run_full_inference(
                 print(f"Successfully SAVED resulting file at {output_file_path}")
             except Exception as e:
                 print(f"Error saving file {excel_name}: {e}")
+
+    combine_file_path = os.path.join(output_path, "combined_inference")
+    os.makedirs(combine_file_path, exist_ok=True)
+    combine_and_filter_xlsx_files(output_path, combine_file_path)
+    print(
+        "Successfully SAVED combined inference file for KPI DETECTION in ",
+        combine_file_path,
+    )