Upload app.py

app.py

@@ -220,8 +220,22 @@ print(f"Model ready on {device}")
 enc = tiktoken.get_encoding('gpt2')
 
 
-def generate_text(prompt, max_new_tokens=100, temperature=0.7, top_k=50, top_p=0.9, repetition_penalty=1.1):
-    """Generate text from prompt with improved sampling"""
+def generate_text(prompt, max_new_tokens=100, temperature=0.6, top_k=40, top_p=0.8, repetition_penalty=1.15):
+    """
+    Generate text from prompt with improved sampling.
+    
+    Note on Grammar Quality:
+    Despite training, grammar may not be perfect due to:
+    1. Model size (124M) - relatively small for perfect grammar
+    2. Training loss measures token prediction, not grammatical correctness
+    3. Shakespeare text has archaic grammar patterns the model learns
+    4. BPE tokenization can split words, making grammar harder to learn
+    
+    For better grammar, try:
+    - Lower temperature (0.5-0.6) for more focused, coherent text
+    - Higher repetition penalty (1.15-1.2) to reduce repetitive patterns
+    - Lower top_p (0.8-0.85) for more conservative sampling
+    """
     try:
         if not model_loaded:
             return "❌ Error: Model not loaded correctly. Please check that model_checkpoint_final.pt is uploaded to HuggingFace Model Hub (shwethd/gpt2-shakespeare-124m)."
@@ -996,6 +1010,13 @@ with gr.Blocks(title="GPT-2 124M Shakespeare Model") as demo:
     - Model Parameters: 124.44M
     - Training Steps: 1,637
     
+    **📝 Note on Grammar Quality:**
+    The model learns token patterns from Shakespeare text, but grammar may not be perfect due to:
+    - Model size (124M is relatively small for perfect grammar)
+    - Training loss measures token prediction, not grammatical correctness
+    - Shakespeare text has archaic grammar patterns
+    - For better grammar, use lower temperature (0.5-0.6) and higher repetition penalty (1.15-1.2)
+    
     Enter a prompt below to generate Shakespeare-style text!
     
     {"⚠️ **Note:** If you see garbled/random text, the model may not have loaded correctly. Check the logs and ensure the model is uploaded to HuggingFace Model Hub: `shwethd/gpt2-shakespeare-124m`" if not model_loaded else ""}
@@ -1020,33 +1041,33 @@ with gr.Blocks(title="GPT-2 124M Shakespeare Model") as demo:
                 label="Temperature",
                 minimum=0.1,
                 maximum=2.0,
-                value=0.7,
+                value=0.6,
                 step=0.1,
-                info="Lower = more focused, Higher = more creative (0.7 recommended for better coherence)"
+                info="Lower = more focused/grammatical (0.5-0.6 recommended). Higher = more creative but less coherent"
             )
             top_k = gr.Slider(
                 label="Top-K",
                 minimum=10,
                 maximum=100,
-                value=50,
+                value=40,
                 step=10,
-                info="Number of top tokens to consider"
+                info="Number of top tokens to consider. Lower (30-40) = more focused, Higher = more diverse"
             )
             top_p = gr.Slider(
                 label="Top-P (Nucleus)",
                 minimum=0.1,
                 maximum=1.0,
-                value=0.85,
+                value=0.8,
                 step=0.05,
-                info="Nucleus sampling - 0.85-0.9 recommended. Lower (0.3) = too restrictive, Higher (0.95+) = too random"
+                info="Nucleus sampling. Lower (0.75-0.8) = more grammatical/coherent. Higher (0.9+) = more random"
             )
             repetition_penalty = gr.Slider(
                 label="Repetition Penalty",
                 minimum=1.0,
                 maximum=1.5,
-                value=1.1,
+                value=1.15,
                 step=0.05,
-                info="Penalize repeated tokens - higher = less repetition (1.1 recommended)"
+                info="Penalize repeated tokens. Higher (1.15-1.2) = better grammar/less repetition. Lower = more repetitive"
             )
             generate_btn = gr.Button("Generate", variant="primary")
         
@@ -1063,20 +1084,21 @@ with gr.Blocks(title="GPT-2 124M Shakespeare Model") as demo:
     examples = gr.Examples(
         examples=[
             # Format: [prompt, max_tokens, temperature, top_k, top_p, repetition_penalty]
-            ["First Citizen:", 100, 0.7, 50, 0.85, 1.1],
-            ["ROMEO:", 100, 0.65, 45, 0.88, 1.15],  # Romantic - slightly lower temp
-            ["To be or not", 80, 0.6, 40, 0.85, 1.2],  # Quote - more focused
-            ["HAMLET:", 100, 0.7, 50, 0.85, 1.1],
-            ["MACBETH:", 100, 0.7, 50, 0.85, 1.1],
-            ["JULIET:", 100, 0.65, 45, 0.88, 1.15],  # Romantic
-            ["KING:", 100, 0.7, 50, 0.85, 1.1],
-            ["LADY MACBETH:", 100, 0.7, 50, 0.85, 1.1],
-            ["OTHELLO:", 100, 0.7, 50, 0.85, 1.1],
-            ["What light through yonder", 100, 0.65, 45, 0.88, 1.15],  # Romantic quote
-            ["All the world's a stage", 100, 0.7, 50, 0.85, 1.1],  # Metaphorical
-            ["Double, double toil and trouble", 80, 0.7, 50, 0.85, 1.15],  # Witches chant
-            ["Friends, Romans, countrymen", 100, 0.7, 50, 0.85, 1.1],  # Speech
-            ["A rose by any other name", 100, 0.65, 45, 0.88, 1.15],  # Romantic quote
+            # Optimized for better grammar: lower temp (0.5-0.6), lower top_p (0.75-0.8), higher rep_penalty (1.15-1.2)
+            ["First Citizen:", 100, 0.6, 40, 0.8, 1.15],
+            ["ROMEO:", 100, 0.55, 35, 0.78, 1.2],  # Romantic - lower temp for coherence
+            ["To be or not", 80, 0.5, 30, 0.75, 1.2],  # Quote - very focused for grammar
+            ["HAMLET:", 100, 0.6, 40, 0.8, 1.15],
+            ["MACBETH:", 100, 0.6, 40, 0.8, 1.15],
+            ["JULIET:", 100, 0.55, 35, 0.78, 1.2],  # Romantic - lower temp
+            ["KING:", 100, 0.6, 40, 0.8, 1.15],
+            ["LADY MACBETH:", 100, 0.6, 40, 0.8, 1.15],
+            ["OTHELLO:", 100, 0.6, 40, 0.8, 1.15],
+            ["What light through yonder", 100, 0.55, 35, 0.78, 1.2],  # Romantic quote
+            ["All the world's a stage", 100, 0.6, 40, 0.8, 1.15],  # Metaphorical
+            ["Double, double toil and trouble", 80, 0.6, 40, 0.8, 1.15],  # Witches chant
+            ["Friends, Romans, countrymen", 100, 0.6, 40, 0.8, 1.15],  # Speech
+            ["A rose by any other name", 100, 0.55, 35, 0.78, 1.2],  # Romantic quote
         ],
         inputs=[prompt_input, max_tokens, temperature, top_k, top_p, repetition_penalty]
     )