Update app.py

app.py

@@ -132,10 +132,7 @@ MODEL_PATH = "mistralai/Mistral-7B-Instruct-v0.3"
 #MODEL_PATH = "tiiuae/falcon-rw-1b"
 
 bnb_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_compute_dtype=torch.float16,
-    bnb_4bit_use_double_quant=True,
-    bnb_4bit_quant_type="nf4"
+    load_in_8bit=True,
 )
 
 mistral_tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH,use_auth_token=True)
@@ -147,6 +144,8 @@ judge_llm = AutoModelForCausalLM.from_pretrained(
     use_auth_token=True
 )
 
+print(judge_llm.hf_device_map)
+
 judge_pipeline = pipeline(
     "text-generation",
     model=judge_llm,
@@ -154,10 +153,14 @@ judge_pipeline = pipeline(
     max_new_tokens=128,
     temperature=0.3,
     top_p=0.9,
-    do_sample=True,
+    do_sample=False,
     repetition_penalty=1.1,
 )
 
+output = judge_pipeline("Q: What is Python?\nA:", max_new_tokens=128)[0]['generated_text']
+print(output)
+
+
 
 # embedding model
 from sentence_transformers import SentenceTransformer
@@ -1700,11 +1703,24 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
             skills=data["skills"], seniority=data["seniority"], difficulty_adjustment=None,
             voice_label="neutral", face_label="neutral"
         )
+        #here the original one 
+        # first_q = groq_llm.predict(prompt)
+        # # Evaluate Q for quality
+        # q_eval = eval_question_quality(first_q, data["job_role"], data["seniority"], None)
+        # state["questions"].append(first_q)
+        # state["question_evaluations"].append(q_eval)
+
+        #here the testing one
         first_q = groq_llm.predict(prompt)
-        # Evaluate Q for quality
-        q_eval = eval_question_quality(first_q, data["job_role"], data["seniority"], None)
+        q_eval = {
+            "Score": "N/A",
+            "Reasoning": "Skipped to reduce processing time",
+            "Improvements": []
+        }
         state["questions"].append(first_q)
         state["question_evaluations"].append(q_eval)
+
+        
         state["conversation_history"].append({'role': 'Interviewer', 'content': first_q})
         audio_path = bark_tts(first_q)
         # LOG
@@ -1725,11 +1741,19 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         state["conversation_history"].append({'role': 'Candidate', 'content': transcript})
 
         # --- 1. Emotion analysis ---
-        voice_label = analyze_audio_emotion(audio_path)
-        face_label = analyze_video_emotions(video_path)
+        # voice_label = analyze_audio_emotion(audio_path)
+        # face_label = analyze_video_emotions(video_path)
+        # state["voice_labels"].append(voice_label)
+        # state["face_labels"].append(face_label)
+
+        #just for testing
+        voice_label = "neutral"
+        face_label = "neutral"
         state["voice_labels"].append(voice_label)
         state["face_labels"].append(face_label)
 
+        
+
         # --- 2. Evaluate previous Q and Answer ---
         last_q = state["questions"][-1]
         q_eval = state["question_evaluations"][-1]  # Already in state
@@ -1747,9 +1771,14 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
             state["difficulty_adjustment"] = None
 
         # --- 4. Effective confidence ---
-        eff_conf = interpret_confidence(voice_label, face_label, answer_score)
+        # eff_conf = interpret_confidence(voice_label, face_label, answer_score)
+        # state["effective_confidences"].append(eff_conf)
+
+        #just for testing:
+        eff_conf = {"effective_confidence": 0.6}
         state["effective_confidences"].append(eff_conf)
 
+
         # --- LOG ---
         state["log"].append({
             "type": "answer",
@@ -1779,8 +1808,9 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
                             f"- *Answer*: {state['answers'][i]}\n"
                             f"- *Q Eval*: {state['question_evaluations'][i]}\n"
                             f"- *A Eval*: {state['answer_evaluations'][i]}\n"
-                            f"- *Face Emotion: {state['face_labels'][i]}, **Voice Emotion*: {state['voice_labels'][i]}\n"
-                            f"- *Effective Confidence*: {state['effective_confidences'][i]['effective_confidence']}\n"
+                            #also this are removed just for testing :(
+                            # f"- *Face Emotion: {state['face_labels'][i]}, **Voice Emotion*: {state['voice_labels'][i]}\n"
+                            # f"- *Effective Confidence*: {state['effective_confidences'][i]['effective_confidence']}\n"
                             f"- *Time*: {state['timings'][i]}s\n")
             summary += f"\n\n⏺ Full log saved as {log_file}."
             return (state, gr.update(visible=True, value=summary), gr.update(value=None), gr.update(value=None), gr.update(value=None), gr.update(value=None), gr.update(visible=True, value=f"Last Detected — Face: {face_label}, Voice: {voice_label}"))