Update app.py

app.py

@@ -1,16 +1,16 @@
-import gradio as gr
-from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
 import torch
-from sentence_transformers import SentenceTransformer, util
-import numpy as np
+from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
 
 # Configuração de dispositivo
 DEVICE = 0 if torch.cuda.is_available() else -1
 TORCH_DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32
 
-# Modelo A: Falcon RW 1B
-model_a = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-rw-1b", torch_dtype=TORCH_DTYPE)
-tokenizer_a = AutoTokenizer.from_pretrained("tiiuae/falcon-rw-1b")
+# Carrega modelo A - LLaMA 2
+model_a = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-2-7b-chat-hf",
+    torch_dtype=TORCH_DTYPE
+)
+tokenizer_a = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
 
 pipe_a = pipeline(
     "text-generation",
@@ -21,9 +21,12 @@ pipe_a = pipeline(
     pad_token_id=tokenizer_a.eos_token_id
 )
 
-# Modelo B: Mistral 7B Instruct
-model_b = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1", torch_dtype=TORCH_DTYPE)
-tokenizer_b = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
+# Carrega modelo B - Falcon 7B Instruct (sem autenticação)
+model_b = AutoModelForCausalLM.from_pretrained(
+    "tiiuae/falcon-7b-instruct",
+    torch_dtype=TORCH_DTYPE
+)
+tokenizer_b = AutoTokenizer.from_pretrained("tiiuae/falcon-7b-instruct")
 
 pipe_b = pipeline(
     "text-generation",
@@ -34,124 +37,24 @@ pipe_b = pipeline(
     pad_token_id=tokenizer_b.eos_token_id
 )
 
-# Classificador de sentimento
-sentiment_arbiter = pipeline(
-    "text-classification",
-    model="nlptown/bert-base-multilingual-uncased-sentiment",
-    device=DEVICE
-)
-
-# Modelo de similaridade semântica
-similarity_model = SentenceTransformer(
-    "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
-    device="cuda" if torch.cuda.is_available() else "cpu"
-)
-
-def semantic_similarity(text1, text2):
-    if not text1.strip() or not text2.strip():
-        return 0.0
-    embeddings = similarity_model.encode([text1, text2], convert_to_tensor=True)
-    cosine_scores = util.pytorch_cos_sim(embeddings[0], embeddings[1])
-    return cosine_scores.item()
-
-def format_mistral_prompt(user_input):
-    return f"<s>[INST] {user_input.strip()} [/INST]"
-
-def judge_response(question, response_a, response_b):
-    sentiment_a = sentiment_arbiter(response_a)[0]
-    sentiment_b = sentiment_arbiter(response_b)[0]
-
-    score_sent_a = int(sentiment_a['label'][0])
-    score_sent_b = int(sentiment_b['label'][0])
-
-    sim_score_a = semantic_similarity(question, response_a)
-    sim_score_b = semantic_similarity(question, response_b)
-
-    conciseness_a = min(1.0, 50 / max(1, len(response_a.split())))
-    conciseness_b = min(1.0, 50 / max(1, len(response_b.split())))
-
-    WEIGHT_SENTIMENT = 0.4
-    WEIGHT_RELEVANCE = 0.5
-    WEIGHT_CONCISENESS = 0.1
-
-    total_a = (WEIGHT_SENTIMENT * score_sent_a +
-               WEIGHT_RELEVANCE * sim_score_a +
-               WEIGHT_CONCISENESS * conciseness_a)
-
-    total_b = (WEIGHT_SENTIMENT * score_sent_b +
-               WEIGHT_RELEVANCE * sim_score_b +
-               WEIGHT_CONCISENESS * conciseness_b)
-
-    THRESHOLD = 0.15
-    if abs(total_a - total_b) < THRESHOLD:
-        winner = "Modelo A" if score_sent_a >= score_sent_b else "Modelo B"
-        final_response = response_a if score_sent_a >= score_sent_b else response_b
-    else:
-        winner = "Modelo A" if total_a > total_b else "Modelo B"
-        final_response = response_a if total_a > total_b else response_b
-
-    print(f"\nA: S:{score_sent_a:.2f} R:{sim_score_a:.2f} C:{conciseness_a:.2f} T:{total_a:.2f}")
-    print(f"B: S:{score_sent_b:.2f} R:{sim_score_b:.2f} C:{conciseness_b:.2f} T:{total_b:.2f}")
-    print(f"Vencedor: {winner} Diferença: {abs(total_a - total_b):.2f}")
-
-    return winner, final_response
-
-def chatbot(prompt):
-    prompt_pt = "Responda em português: " + prompt
-    mistral_prompt = format_mistral_prompt(prompt_pt)
-
-    response_a = pipe_a(
-        prompt_pt,
-        max_new_tokens=60,
-        temperature=0.7,
-        top_k=50,
-        top_p=0.9,
-        repetition_penalty=1.2,
-    )[0]['generated_text'].strip()
-
-    response_b = pipe_b(
-        mistral_prompt,
-        max_new_tokens=60,
-        temperature=0.7,
-        top_k=50,
-        top_p=0.9,
-        repetition_penalty=1.2,
-    )[0]['generated_text'].strip()
-
-    winner, final_response = judge_response(prompt, response_a, response_b)
-    return prompt, response_a, response_b, winner, final_response
-
-css = """
-footer {visibility: hidden}
-.output-text {font-size: 16px !important}
-"""
-
-with gr.Blocks(css=css) as demo:
-    gr.Markdown("# 🤖 Chatbot com Julgamento Aprimorado")
-    gr.Markdown("Compara respostas de dois modelos usando múltiplos critérios de qualidade")
-
-    with gr.Row():
-        inp = gr.Textbox(label="Digite sua pergunta:", lines=2, placeholder="Escreva sua pergunta em português...")
-        btn = gr.Button("Enviar")
+# Funções auxiliares para formatar o prompt
+def format_llama_prompt(user_input):
+    return f"[INST] <<SYS>>\nVocê é um assistente útil.\n<</SYS>>\n\n{user_input.strip()} [/INST]"
 
-    with gr.Row():
-        with gr.Column():
-            gr.Markdown("### Modelo A (Falcon RW 1B)")
-            out_a = gr.Textbox(label="Resposta", interactive=False)
-        with gr.Column():
-            gr.Markdown("### Modelo B (Mistral 7B Instruct)")
-            out_b = gr.Textbox(label="Resposta", interactive=False)
+def format_falcon_prompt(user_input):
+    return f"Responda em português: {user_input.strip()}"
 
-    with gr.Row():
-        with gr.Column(scale=2):
-            winner_out = gr.Textbox(label="🏆 Modelo Vencedor", interactive=False)
-        with gr.Column(scale=3):
-            final_out = gr.Textbox(label="💡 Resposta Escolhida", interactive=False)
+# Interface simples para testar os modelos
+if __name__ == "__main__":
+    while True:
+        prompt = input("\nDigite uma pergunta (ou 'sair'): ").strip()
+        if prompt.lower() == "sair":
+            break
 
-    btn.click(
-        fn=chatbot,
-        inputs=inp,
-        outputs=[inp, out_a, out_b, winner_out, final_out]
-    )
+        print("\n=== Resposta do LLaMA 2 ===")
+        llama_response = pipe_a(format_llama_prompt(prompt), max_new_tokens=200)[0]['generated_text']
+        print(llama_response)
 
-demo.launch()
+        print("\n=== Resposta do Falcon 7B ===")
+        falcon_response = pipe_b(format_falcon_prompt(prompt), max_new_tokens=200)[0]['generated_text']
+        print(falcon_response)