Upload code/train_sft.py with huggingface_hub

code/train_sft.py

@@ -290,10 +290,10 @@ def ensure_cloud_data_and_checkpoint(version: str, data_dir: str, ckpt_path: str
 
 
 # ------------------------------------------------------------------------------
-# 7. Live Generation Helper for Training Telemetry
+# 7. Live Generation Helper for Training Telemetry (Multi-Domain Previews)
 # ------------------------------------------------------------------------------
 @torch.no_grad()
-def generate_sample_preview(model, tokenizer, device, prompt="<|user|>\nWho created you?<|endofturn|>\n<|assistant|>\n", max_new_tokens=40):
+def generate_sample_preview(model, tokenizer, device, prompt: str, max_new_tokens=60):
     if tokenizer is None:
         return ""
     model.eval()
@@ -309,6 +309,29 @@ def generate_sample_preview(model, tokenizer, device, prompt="<|user|>\nWho crea
         model.train()
 
 
+def run_live_eval_previews(model, tokenizer, device):
+    """Runs a suite of multi-domain test prompts to visually monitor model progress."""
+    if tokenizer is None:
+        return
+
+    test_suite = [
+        ("Identity",    "<|user|>\nWho created you?<|endofturn|>\n<|assistant|>\n", 40),
+        ("Greeting",    "<|user|>\nHello! Kaise ho aap?<|endofturn|>\n<|assistant|>\n", 40),
+        ("Translation", "<|user|>\nTranslate the following English sentence to Hindi:\n\"Artificial intelligence is shaping the future.\"<|endofturn|>\n<|assistant|>\n", 50),
+        ("Logic & Math", "<|user|>\nIf a train travels at 60 km/h, how far will it travel in 3.5 hours?<|endofturn|>\n<|assistant|>\n", 80),
+        ("Coding",      "<|user|>\nWrite a Python function to check if a number is prime.<|endofturn|>\n<|assistant|>\n", 70)
+    ]
+
+    print("   💬 --- [LIVE MULTI-DOMAIN TEST PREVIEWS] ---")
+    for category, prompt, max_tok in test_suite:
+        answer = generate_sample_preview(model, tokenizer, device, prompt, max_new_tokens=max_tok)
+        # Format response cleanly for terminal
+        clean_ans = answer.replace("\n", " ").strip()
+        if len(clean_ans) > 120:
+            clean_ans = clean_ans[:117] + "..."
+        print(f"      • [{category:12s}]: \"{clean_ans}\"")
+
+
 # ------------------------------------------------------------------------------
 # 8. Main Ultra-Optimized Training Engine
 # ------------------------------------------------------------------------------
@@ -661,10 +684,9 @@ def main():
                     # Print Domain Loss Breakdown
                     print("   📊 Domain Breakdown: " + " | ".join([f"{k[:6]}: {v:.3f}" for k, v in d_losses.items()]))
                     
-                    # Live Sample Generation Preview
+                    # Live Multi-Domain Sample Generation Previews
                     if tokenizer is not None:
-                        preview_id = generate_sample_preview(model, tokenizer, device, "<|user|>\nWho created you?<|endofturn|>\n<|assistant|>\n", 30)
-                        print(f"   💬 [Preview Identity]: \"{preview_id}\"")
+                        run_live_eval_previews(model, tokenizer, device)
                     
                     is_best = v_loss < best_val_loss
                     if is_best: