test_norah.py

from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList
import torch

model_name = r"C:\Users\HP-Victus\GVAIDAL\Norah"  # Use full path

print("🔄 Loading tokenizer and model...")

tokenizer = AutoTokenizer.from_pretrained(model_name, local_files_only=True)
model = AutoModelForCausalLM.from_pretrained(model_name, local_files_only=True, torch_dtype=torch.float16, device_map="auto")

def format_prompt(user_input):
    return (
        "Tu es un assistant IA utile et intelligent qui répond toujours en français avec des réponses courtes et claires.\n\n"
        "Utilisateur: Bonjour, comment vas-tu ?\n"
        "Assistant: Bonjour ! Je vais bien, merci. Comment puis-je vous aider ?\n\n"
        f"Utilisateur: {user_input}\n"
        "Assistant:"
    )

# Test conversation
prompt = format_prompt("Bonjour, comment puis-je vous aider aujourd'hui ?")

inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512).to("cuda" if torch.cuda.is_available() else "cpu")

print("📝 Generating response...")

# Encode names to block
bad_words = ["Jean", "Marie", "Bouchard", "Pierre", "Louis", "Antoine", "Jacques", "Robert", "Roper"]
bad_words_ids = [tokenizer.encode(word, add_special_tokens=False) for word in bad_words]

# Stopping criteria: Stop at sentence completion
class StopOnSentenceEnd(StoppingCriteria):
    def __call__(self, input_ids, scores, **kwargs):
        stop_tokens = [tokenizer.encode(".", add_special_tokens=False)[0],
                       tokenizer.encode("!", add_special_tokens=False)[0],
                       tokenizer.encode("?", add_special_tokens=False)[0]]
        return any(input_ids[0, -1].item() == stop for stop in stop_tokens)

stopping_criteria = StoppingCriteriaList([StopOnSentenceEnd()])

# Generate response
outputs = model.generate(
    **inputs,
    max_length=100,  # Allows complete sentences
    min_length=10,   # Ensures at least some response
    do_sample=True,  # Allows varied responses
    temperature=0.7,  # More natural responses
    top_p=0.9,       # Higher probability for relevant words
    repetition_penalty=1.5,  # Prevents repetition but keeps coherence
    eos_token_id=model.config.eos_token_id,
    stopping_criteria=stopping_criteria  # Ensures sentence completion
)

# Decode and display response
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("💬 Model Response:", response)