Upload test_model.py

test_model.py

@@ -0,0 +1,85 @@
+"""Quick test: load palindrome model, generate samples for multiple themes."""
+import re
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+print(f"Device: {device}")
+
+def _normalize(text):
+    return re.sub(r'[^a-zA-Z0-9]', '', text).lower()
+
+def is_palindrome(text):
+    cleaned = _normalize(text)
+    return cleaned == cleaned[::-1]
+
+# Test V1 (completed)
+model_id = "SantiagoC/palindrome-grpo"
+print(f"\nLoading {model_id}...")
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    torch_dtype=torch.float16,
+    device_map="auto",
+)
+
+themes = [
+    "cars", "nature", "food", "love", 
+    "space", "animals", "music", "destiny",
+    "technology", "water",
+]
+
+print(f"\nModel: {model_id}")
+print("=" * 70)
+for theme in themes:
+    messages = [
+        {"role": "system", "content": "You generate palindromes. A palindrome reads the same forwards and backwards after removing non-alphanumeric characters and lowercasing."},
+        {"role": "user", "content": f"Here are some examples of palindromes:\n- 'A Toyota' (about cars)\n- 'No lemon, no melon' (about food)\n- 'Was it a car or a cat I saw?' (about animals)\n\nNow: Create a phrase that reads the same forwards and backwards, about {theme}."},
+    ]
+    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer(text, return_tensors="pt").to(device)
+    
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=80,
+            temperature=0.9,
+            do_sample=True,
+            pad_token_id=tokenizer.pad_token_id,
+        )
+    
+    completion = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
+    pal = "YES" if is_palindrome(completion) else "NO "
+    cleaned = _normalize(completion)
+    print(f"\nTheme: {theme:15s} | palindrome: {pal} | chars: {len(cleaned):3d}")
+    print(f"  → {completion[:150]}")
+
+# Also test V5 if available
+try:
+    model_id_v5 = "SantiagoC/palindrome-grpo-v5"
+    tokenizer_v5 = AutoTokenizer.from_pretrained(model_id_v5)
+    model_v5 = AutoModelForCausalLM.from_pretrained(model_id_v5, torch_dtype=torch.float16, device_map="auto")
+    if tokenizer_v5.pad_token is None:
+        tokenizer_v5.pad_token = tokenizer_v5.eos_token
+
+    print(f"\n\nModel: {model_id_v5}")
+    print("=" * 70)
+    for theme in themes[:5]:
+        messages = [
+            {"role": "system", "content": "You generate palindromes. A palindrome reads the same forwards and backwards after removing non-alphanumeric characters and lowercasing."},
+            {"role": "user", "content": f"Here are some examples of palindromes:\n- 'A Toyota' (about cars)\n- 'No lemon, no melon' (about food)\n- 'Was it a car or a cat I saw?' (about animals)\n\nNow: Create a phrase that reads the same forwards and backwards, about {theme}."},
+        ]
+        text = tokenizer_v5.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = tokenizer_v5(text, return_tensors="pt").to(device)
+        with torch.no_grad():
+            outputs = model_v5.generate(**inputs, max_new_tokens=80, temperature=0.9, do_sample=True, pad_token_id=tokenizer_v5.pad_token_id)
+        completion = tokenizer_v5.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
+        pal = "YES" if is_palindrome(completion) else "NO "
+        cleaned = _normalize(completion)
+        print(f"\nTheme: {theme:15s} | palindrome: {pal} | chars: {len(cleaned):3d}")
+        print(f"  → {completion[:150]}")
+except Exception as e:
+    print(f"\nV5 model not available yet: {e}")