Spaces:
Paused
Paused
Robin Genolet
commited on
Commit
·
5fd44e9
1
Parent(s):
dad4228
fix: params
Browse files- app.py +16 -90
- utils/epfl_meditron_utils.py +7 -3
app.py
CHANGED
@@ -10,6 +10,8 @@ import subprocess
|
|
10 |
import sys
|
11 |
import io
|
12 |
|
|
|
|
|
13 |
from utils.default_values import get_system_prompt, get_guidelines_dict
|
14 |
from utils.epfl_meditron_utils import get_llm_response, gptq_model_options
|
15 |
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
|
|
17 |
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
|
18 |
from sklearn.metrics import classification_report
|
19 |
|
20 |
-
DATA_FOLDER = "data/"
|
21 |
-
|
22 |
POC_VERSION = "0.1.0"
|
23 |
-
MAX_QUESTIONS = 10
|
24 |
-
AVAILABLE_LANGUAGES = ["DE", "EN", "FR"]
|
25 |
|
26 |
st.set_page_config(page_title='Medgate Whisper PoC', page_icon='public/medgate.png')
|
27 |
|
28 |
-
# Azure apparently truncates message if longer than 200, see
|
29 |
-
MAX_SYSTEM_MESSAGE_TOKENS = 200
|
30 |
-
|
31 |
-
|
32 |
-
def format_question(q):
|
33 |
-
res = q
|
34 |
-
|
35 |
-
# Remove numerical prefixes, if any, e.g. '1. [...]'
|
36 |
-
if re.match(r'^[0-9].\s', q):
|
37 |
-
res = res[3:]
|
38 |
-
|
39 |
-
# Replace doc reference by doc name
|
40 |
-
if len(st.session_state["citations"]) > 0:
|
41 |
-
for source_ref in re.findall(r'\[doc[0-9]+\]', res):
|
42 |
-
citation_number = int(re.findall(r'[0-9]+', source_ref)[0])
|
43 |
-
citation_index = citation_number - 1 if citation_number > 0 else 0
|
44 |
-
citation = st.session_state["citations"][citation_index]
|
45 |
-
source_title = citation["title"]
|
46 |
-
res = res.replace(source_ref, '[' + source_title + ']')
|
47 |
-
|
48 |
-
return res.strip()
|
49 |
-
|
50 |
-
|
51 |
-
def get_text_from_row(text):
|
52 |
-
res = str(text)
|
53 |
-
if res == "nan":
|
54 |
-
return ""
|
55 |
-
return res
|
56 |
-
def get_questions_from_df(df, lang, test_scenario_name):
|
57 |
-
questions = []
|
58 |
-
for i, row in df.iterrows():
|
59 |
-
questions.append({
|
60 |
-
"question": row[lang + ": Fragen"],
|
61 |
-
"answer": get_text_from_row(row[test_scenario_name]),
|
62 |
-
"question_id": uuid.uuid4()
|
63 |
-
})
|
64 |
-
return questions
|
65 |
-
|
66 |
-
|
67 |
-
def get_questions(df, lead_symptom, lang, test_scenario_name):
|
68 |
-
print(str(st.session_state["lead_symptom"]) + " -> " + lead_symptom)
|
69 |
-
print(str(st.session_state["scenario_name"]) + " -> " + test_scenario_name)
|
70 |
-
if st.session_state["lead_symptom"] != lead_symptom or st.session_state["scenario_name"] != test_scenario_name:
|
71 |
-
st.session_state["lead_symptom"] = lead_symptom
|
72 |
-
st.session_state["scenario_name"] = test_scenario_name
|
73 |
-
symptom_col_name = st.session_state["language"] + ": Symptome"
|
74 |
-
df_questions = df[(df[symptom_col_name] == lead_symptom)]
|
75 |
-
st.session_state["questions"] = get_questions_from_df(df_questions, lang, test_scenario_name)
|
76 |
-
|
77 |
-
return st.session_state["questions"]
|
78 |
-
|
79 |
-
|
80 |
def display_streamlit_sidebar():
|
81 |
st.sidebar.title("Local LLM PoC " + str(POC_VERSION))
|
82 |
|
83 |
st.sidebar.write('**Parameters**')
|
84 |
form = st.sidebar.form("config_form", clear_on_submit=True)
|
85 |
|
86 |
-
model_name_or_path = form.selectbox("Select model", gptq_model_options())
|
|
|
87 |
|
88 |
temperature = form.slider(label="Temperature", min_value=0.0, max_value=1.0, step=0.01, value=st.session_state["temperature"])
|
89 |
do_sample = form.checkbox('do_sample', value=st.session_state["do_sample"])
|
@@ -98,6 +45,15 @@ def display_streamlit_sidebar():
|
|
98 |
st.session_state['session_started'] = True
|
99 |
|
100 |
st.session_state["session_events"] = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
st.session_state["model_name_or_path"] = model_name_or_path
|
102 |
st.session_state["temperature"] = temperature
|
103 |
st.session_state["do_sample"] = do_sample
|
@@ -123,10 +79,7 @@ def init_session_state():
|
|
123 |
st.session_state["system_prompt"] = "You are a medical expert that provides answers for a medically trained audience"
|
124 |
st.session_state["prompt"] = ""
|
125 |
st.session_state["llm_messages"] = []
|
126 |
-
|
127 |
-
def get_genders():
|
128 |
-
return ['Male', 'Female']
|
129 |
-
|
130 |
|
131 |
def display_session_overview():
|
132 |
st.subheader('History of LLM queries')
|
@@ -156,33 +109,6 @@ def display_session_overview():
|
|
156 |
st.write("Total compute time (ms): " + str(total_time))
|
157 |
|
158 |
|
159 |
-
def plot_report(title, expected, predicted, display_labels):
|
160 |
-
st.markdown('#### ' + title)
|
161 |
-
conf_matrix = confusion_matrix(expected, predicted, labels=display_labels)
|
162 |
-
conf_matrix_plot = ConfusionMatrixDisplay(confusion_matrix=conf_matrix, display_labels=display_labels)
|
163 |
-
conf_matrix_plot.plot()
|
164 |
-
st.pyplot(plt.gcf())
|
165 |
-
|
166 |
-
report = classification_report(expected, predicted, output_dict=True)
|
167 |
-
df_report = pd.DataFrame(report).transpose()
|
168 |
-
st.write(df_report)
|
169 |
-
|
170 |
-
df_rp = df_report
|
171 |
-
df_rp = df_rp.drop('support', axis=1)
|
172 |
-
df_rp = df_rp.drop(['accuracy', 'macro avg', 'weighted avg'])
|
173 |
-
|
174 |
-
try:
|
175 |
-
ax = df_rp.plot(kind="bar", legend=True)
|
176 |
-
for container in ax.containers:
|
177 |
-
ax.bar_label(container, fontsize=7)
|
178 |
-
plt.xticks(rotation=45)
|
179 |
-
plt.legend(loc=(1.04, 0))
|
180 |
-
st.pyplot(plt.gcf())
|
181 |
-
except Exception as e:
|
182 |
-
# Out of bounds
|
183 |
-
pass
|
184 |
-
|
185 |
-
|
186 |
def get_prompt_format(model_name):
|
187 |
formatted_text = ""
|
188 |
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):
|
|
202 |
|
203 |
'''
|
204 |
|
205 |
-
return
|
206 |
|
207 |
def format_prompt(template, system_message, prompt):
|
208 |
if template == "":
|
@@ -227,7 +153,7 @@ def display_llm_output():
|
|
227 |
formatted_prompt = format_prompt(prompt_format, system_prompt, prompt)
|
228 |
print(f"Formatted prompt: {format_prompt}")
|
229 |
llm_response = get_llm_response(
|
230 |
-
st.session_state["
|
231 |
st.session_state["temperature"],
|
232 |
st.session_state["do_sample"],
|
233 |
st.session_state["top_p"],
|
|
|
10 |
import sys
|
11 |
import io
|
12 |
|
13 |
+
import inspect
|
14 |
+
|
15 |
from utils.default_values import get_system_prompt, get_guidelines_dict
|
16 |
from utils.epfl_meditron_utils import get_llm_response, gptq_model_options
|
17 |
from utils.openai_utils import get_available_engines, get_search_query_type_options
|
|
|
19 |
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
|
20 |
from sklearn.metrics import classification_report
|
21 |
|
|
|
|
|
22 |
POC_VERSION = "0.1.0"
|
|
|
|
|
23 |
|
24 |
st.set_page_config(page_title='Medgate Whisper PoC', page_icon='public/medgate.png')
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
def display_streamlit_sidebar():
|
27 |
st.sidebar.title("Local LLM PoC " + str(POC_VERSION))
|
28 |
|
29 |
st.sidebar.write('**Parameters**')
|
30 |
form = st.sidebar.form("config_form", clear_on_submit=True)
|
31 |
|
32 |
+
model_name_or_path = form.selectbox("Select model", gptq_model_options(), value=st.session_state["model_index"])
|
33 |
+
model_name_or_path_other = form.text_input('Or input any GPTQ model', value=st.session_state["model_name_or_path_other"])
|
34 |
|
35 |
temperature = form.slider(label="Temperature", min_value=0.0, max_value=1.0, step=0.01, value=st.session_state["temperature"])
|
36 |
do_sample = form.checkbox('do_sample', value=st.session_state["do_sample"])
|
|
|
45 |
st.session_state['session_started'] = True
|
46 |
|
47 |
st.session_state["session_events"] = []
|
48 |
+
|
49 |
+
if len(model_name_or_path_other) > 0:
|
50 |
+
st.session_state["model_name"] = model_name_or_path_other
|
51 |
+
st.session_state["model_name_or_path_other"] = model_name_or_path_other
|
52 |
+
else:
|
53 |
+
st.session_state["model_name"] = model_name_or_path
|
54 |
+
st.session_state["model_index"] = gptq_model_options().index(model_name_or_path)
|
55 |
+
|
56 |
+
|
57 |
st.session_state["model_name_or_path"] = model_name_or_path
|
58 |
st.session_state["temperature"] = temperature
|
59 |
st.session_state["do_sample"] = do_sample
|
|
|
79 |
st.session_state["system_prompt"] = "You are a medical expert that provides answers for a medically trained audience"
|
80 |
st.session_state["prompt"] = ""
|
81 |
st.session_state["llm_messages"] = []
|
82 |
+
|
|
|
|
|
|
|
83 |
|
84 |
def display_session_overview():
|
85 |
st.subheader('History of LLM queries')
|
|
|
109 |
st.write("Total compute time (ms): " + str(total_time))
|
110 |
|
111 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
def get_prompt_format(model_name):
|
113 |
formatted_text = ""
|
114 |
if model_name == "TheBloke/Llama-2-13B-chat-GPTQ" or model_name== "TheBloke/Llama-2-7B-Chat-GPTQ":
|
|
|
128 |
|
129 |
'''
|
130 |
|
131 |
+
return inspect.cleandoc(formatted_text)
|
132 |
|
133 |
def format_prompt(template, system_message, prompt):
|
134 |
if template == "":
|
|
|
153 |
formatted_prompt = format_prompt(prompt_format, system_prompt, prompt)
|
154 |
print(f"Formatted prompt: {format_prompt}")
|
155 |
llm_response = get_llm_response(
|
156 |
+
st.session_state["model_name"],
|
157 |
st.session_state["temperature"],
|
158 |
st.session_state["do_sample"],
|
159 |
st.session_state["top_p"],
|
utils/epfl_meditron_utils.py
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
|
|
2 |
|
3 |
def gptq_model_options():
|
4 |
return [
|
@@ -19,6 +20,8 @@ def get_llm_response(model_name_or_path, temperature, do_sample, top_p, top_k, m
|
|
19 |
print("Formatted prompt:")
|
20 |
print(formatted_prompt)
|
21 |
|
|
|
|
|
22 |
#print("\n\n*** Generate:")
|
23 |
#input_ids = tokenizer(formatted_prompt, return_tensors='pt').input_ids.cuda()
|
24 |
#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
|
|
37 |
repetition_penalty=repetition_penalty
|
38 |
)
|
39 |
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
1 |
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
|
2 |
+
import streamlit as st
|
3 |
|
4 |
def gptq_model_options():
|
5 |
return [
|
|
|
20 |
print("Formatted prompt:")
|
21 |
print(formatted_prompt)
|
22 |
|
23 |
+
st.session_state["llm_messages"].append(formatted_prompt)
|
24 |
+
|
25 |
#print("\n\n*** Generate:")
|
26 |
#input_ids = tokenizer(formatted_prompt, return_tensors='pt').input_ids.cuda()
|
27 |
#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)
|
|
|
40 |
repetition_penalty=repetition_penalty
|
41 |
)
|
42 |
|
43 |
+
pipe_response = pipe(formatted_prompt)
|
44 |
+
st.session_state["llm_messages"].append(pipe_response)
|
45 |
+
print(pipe_response)
|
46 |
+
return pipe_response[0]['generated_text']
|