fix: params

app.py

@@ -10,6 +10,8 @@ import subprocess
 import sys
 import io
 
+import inspect
+
 from utils.default_values import get_system_prompt, get_guidelines_dict
 from utils.epfl_meditron_utils import get_llm_response, gptq_model_options
 from utils.openai_utils import get_available_engines, get_search_query_type_options
@@ -17,73 +19,18 @@ from utils.openai_utils import get_available_engines, get_search_query_type_opti
 from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
 from sklearn.metrics import classification_report
 
-DATA_FOLDER = "data/"
-
 POC_VERSION = "0.1.0"
-MAX_QUESTIONS = 10
-AVAILABLE_LANGUAGES = ["DE", "EN", "FR"]
 
 st.set_page_config(page_title='Medgate Whisper PoC', page_icon='public/medgate.png')
 
-# Azure apparently truncates message if longer than 200, see
-MAX_SYSTEM_MESSAGE_TOKENS = 200
-
-
-def format_question(q):
-    res = q
-
-    # Remove numerical prefixes, if any, e.g. '1. [...]'
-    if re.match(r'^[0-9].\s', q):
-        res = res[3:]
-
-    # Replace doc reference by doc name
-    if len(st.session_state["citations"]) > 0:
-        for source_ref in re.findall(r'\[doc[0-9]+\]', res):
-            citation_number = int(re.findall(r'[0-9]+', source_ref)[0])
-            citation_index = citation_number - 1 if citation_number > 0 else 0
-            citation = st.session_state["citations"][citation_index]
-            source_title = citation["title"]
-            res = res.replace(source_ref, '[' + source_title + ']')
-
-    return res.strip()
-
-
-def get_text_from_row(text):
-    res = str(text)
-    if res == "nan":
-        return ""
-    return res
-def get_questions_from_df(df, lang, test_scenario_name):
-    questions = []
-    for i, row in df.iterrows():
-        questions.append({
-            "question": row[lang + ": Fragen"],
-            "answer": get_text_from_row(row[test_scenario_name]),
-            "question_id": uuid.uuid4()
-        })
-    return questions
-
-
-def get_questions(df, lead_symptom, lang, test_scenario_name):
-    print(str(st.session_state["lead_symptom"]) + " -> " + lead_symptom)
-    print(str(st.session_state["scenario_name"]) + " -> " + test_scenario_name)
-    if st.session_state["lead_symptom"] != lead_symptom or st.session_state["scenario_name"] != test_scenario_name:
-        st.session_state["lead_symptom"] = lead_symptom
-        st.session_state["scenario_name"] = test_scenario_name
-        symptom_col_name = st.session_state["language"] + ": Symptome"
-        df_questions = df[(df[symptom_col_name] == lead_symptom)]
-        st.session_state["questions"] = get_questions_from_df(df_questions, lang, test_scenario_name)
-
-    return st.session_state["questions"]
-
-
 def display_streamlit_sidebar():
     st.sidebar.title("Local LLM PoC " + str(POC_VERSION))
 
     st.sidebar.write('**Parameters**')
     form = st.sidebar.form("config_form", clear_on_submit=True)
 
-    model_name_or_path = form.selectbox("Select model", gptq_model_options())
+    model_name_or_path = form.selectbox("Select model", gptq_model_options(), value=st.session_state["model_index"])
+    model_name_or_path_other = form.text_input('Or input any GPTQ model', value=st.session_state["model_name_or_path_other"])
     
     temperature = form.slider(label="Temperature", min_value=0.0, max_value=1.0, step=0.01, value=st.session_state["temperature"])
     do_sample = form.checkbox('do_sample', value=st.session_state["do_sample"])
@@ -98,6 +45,15 @@ def display_streamlit_sidebar():
         st.session_state['session_started'] = True
 
         st.session_state["session_events"] = []
+        
+        if len(model_name_or_path_other) > 0:
+            st.session_state["model_name"] = model_name_or_path_other
+            st.session_state["model_name_or_path_other"] = model_name_or_path_other
+        else:
+            st.session_state["model_name"] = model_name_or_path
+            st.session_state["model_index"] = gptq_model_options().index(model_name_or_path)
+            
+            
         st.session_state["model_name_or_path"] = model_name_or_path
         st.session_state["temperature"] = temperature
         st.session_state["do_sample"] = do_sample
@@ -123,10 +79,7 @@ def init_session_state():
     st.session_state["system_prompt"] = "You are a medical expert that provides answers for a medically trained audience"
     st.session_state["prompt"] = ""
     st.session_state["llm_messages"] = []
-
-def get_genders():
-    return ['Male', 'Female']
-
+    
 
 def display_session_overview():
     st.subheader('History of LLM queries')
@@ -156,33 +109,6 @@ def display_session_overview():
         st.write("Total compute time (ms): " + str(total_time))
 
 
-def plot_report(title, expected, predicted, display_labels):
-    st.markdown('#### ' + title)
-    conf_matrix = confusion_matrix(expected, predicted, labels=display_labels)
-    conf_matrix_plot = ConfusionMatrixDisplay(confusion_matrix=conf_matrix, display_labels=display_labels)
-    conf_matrix_plot.plot()
-    st.pyplot(plt.gcf())
-
-    report = classification_report(expected, predicted, output_dict=True)
-    df_report = pd.DataFrame(report).transpose()
-    st.write(df_report)
-
-    df_rp = df_report
-    df_rp = df_rp.drop('support', axis=1)
-    df_rp = df_rp.drop(['accuracy', 'macro avg', 'weighted avg'])
-
-    try:
-        ax = df_rp.plot(kind="bar", legend=True)
-        for container in ax.containers:
-            ax.bar_label(container, fontsize=7)
-        plt.xticks(rotation=45)
-        plt.legend(loc=(1.04, 0))
-        st.pyplot(plt.gcf())
-    except Exception as e:
-        # Out of bounds
-        pass
-
-
 def get_prompt_format(model_name):
     formatted_text = ""
     if model_name == "TheBloke/Llama-2-13B-chat-GPTQ" or model_name== "TheBloke/Llama-2-7B-Chat-GPTQ":
@@ -202,7 +128,7 @@ def get_prompt_format(model_name):
                 
                 '''
     
-    return formatted_text.replace("\t", "")
+    return inspect.cleandoc(formatted_text)
 
 def format_prompt(template, system_message, prompt):
     if template == "":
@@ -227,7 +153,7 @@ def display_llm_output():
         formatted_prompt = format_prompt(prompt_format, system_prompt, prompt)
         print(f"Formatted prompt: {format_prompt}")
         llm_response = get_llm_response(
-            st.session_state["model_name_or_path"],
+            st.session_state["model_name"],
             st.session_state["temperature"],
             st.session_state["do_sample"],
             st.session_state["top_p"],

utils/epfl_meditron_utils.py

@@ -1,4 +1,5 @@
 from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
+import streamlit as st
 
 def gptq_model_options():
     return [
@@ -19,6 +20,8 @@ def get_llm_response(model_name_or_path, temperature, do_sample, top_p, top_k, m
     print("Formatted prompt:")
     print(formatted_prompt)
     
+    st.session_state["llm_messages"].append(formatted_prompt)
+    
     #print("\n\n*** Generate:")
     #input_ids = tokenizer(formatted_prompt, return_tensors='pt').input_ids.cuda()
     #output = model.generate(inputs=input_ids, temperature=temperature, do_sample=do_sample, top_p=top_p, top_k=top_k, max_new_tokens=max_new_tokens)
@@ -37,6 +40,7 @@ def get_llm_response(model_name_or_path, temperature, do_sample, top_p, top_k, m
         repetition_penalty=repetition_penalty
     )
 
-    response = pipe(formatted_prompt)[0]['generated_text']
-    print(response)
-    return response
+    pipe_response = pipe(formatted_prompt)
+    st.session_state["llm_messages"].append(pipe_response)
+    print(pipe_response)
+    return pipe_response[0]['generated_text']