feat: add TrainerCallback to stream live training logs to UI

- Implemented GradioLoggerCallback to forward Hugging Face Trainer logs to Gradio
- Replaced pre-loop simulated logging with true per-step feedback
- UI now shows step-by-step progress without freezing or blocking

train_abuse_model.py

@@ -7,6 +7,7 @@ import os
 import time
 import gradio as gr  # ✅ required for progress bar
 from pathlib import Path
+import queue
 
 # Python standard + ML packages
 import pandas as pd
@@ -23,6 +24,7 @@ from huggingface_hub import hf_hub_download
 # Hugging Face transformers
 import transformers
 from transformers import (
+    TrainerCallback,
     AutoTokenizer,
     DebertaV2Tokenizer,
     BertTokenizer, 
@@ -66,6 +68,15 @@ logger.info(f"Transformers version: {transformers.__version__}")
 logger.info("torch.cuda.is_available(): %s", torch.cuda.is_available())
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
+class GradioLoggerCallback(TrainerCallback):
+    def __init__(self, gr_queue):
+        self.gr_queue = gr_queue
+
+    def on_log(self, args, state, control, logs=None, **kwargs):
+        if logs:
+            msg = f"📊 Step {state.global_step}: {logs}"
+            logger.info(msg)
+            self.gr_queue.put(msg)
 
 def evaluate_model_with_thresholds(trainer, test_dataset):
     """Run full evaluation with automatic threshold tuning."""
@@ -191,6 +202,7 @@ train_texts, val_texts, train_labels, val_labels = train_test_split(
 model_name = "microsoft/deberta-v3-base"
 
 def run_training(progress=gr.Progress(track_tqdm=True)):
+    log_queue = queue.Queue()
     if os.path.exists("saved_model/"):
         yield "✅ Trained model found! Skipping training...\n"
         for line in evaluate_saved_model():
@@ -239,7 +251,8 @@ def run_training(progress=gr.Progress(track_tqdm=True)):
             model=model,
             args=training_args,
             train_dataset=train_dataset,
-            eval_dataset=val_dataset
+            eval_dataset=val_dataset,
+            callbacks=[GradioLoggerCallback(log_queue)]
         )
 
         logger.info("Training started with %d samples", len(train_dataset))        
@@ -262,6 +275,10 @@ def run_training(progress=gr.Progress(track_tqdm=True)):
         # Start training!
         trainer.train()
 
+        # Drain queue to UI
+        while not log_queue.empty():
+            yield log_queue.get()
+        
         progress(1.0)
         yield "✅ Progress: 100%\n"