Skip to content

Instantly share code, notes, and snippets.

@iamaziz
Last active April 18, 2024 00:01
Show Gist options
  • Star 1 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save iamaziz/55ace51054d96ba820f0913c5a320a2d to your computer and use it in GitHub Desktop.
Save iamaziz/55ace51054d96ba820f0913c5a320a2d to your computer and use it in GitHub Desktop.
Ollama UI for Arabic (right-to-left) based on: https://github.com/iamaziz/ollachat (example in the comments)
import os
import json
import datetime
import streamlit as st
from llama_index.llms import Ollama
from llama_index.llms import ChatMessage
# https://docs.llamaindex.ai/en/stable/examples/llm/ollama.html
import ollama
models = ollama.list()["models"]
OLLAMA_MODELS = [model for model in models if "arabic" in model["name"]]
def update_formatting(text):
if is_arabic(text):
return f"<div style='direction: rtl; text-align: right;'>{text}</div>"
return f"{text}"
def st_ollama(model_name, user_question, chat_history_key):
llm_ = Ollama(model=model_name)
if chat_history_key not in st.session_state.keys():
st.session_state[chat_history_key] = []
print_chat_history_timeline(chat_history_key)
# run the model
if user_question:
st.session_state[chat_history_key].append({"content": f"{user_question}", "role": "user"})
with st.chat_message("question", avatar="🧑‍🚀"):
user_question = update_formatting(user_question)
# st.write(user_question)
# change color of user question to blue
# user_question = f"<span style='color: #004d4d;'>{user_question}</span>"
st.markdown(user_question, unsafe_allow_html=True)
messages = [ChatMessage(content=message["content"], role=message["role"]) for message in st.session_state[chat_history_key]]
with st.spinner("افكر كيف ارد 🤔 ... كلمة كلمة "):
response = llm_.stream_chat(messages)
# streaming response
with st.chat_message("response", avatar="🤖"):
chat_box = st.empty()
response_message = ""
for chunk in response:
response_message += chunk.delta
chat_box.markdown(response_message, unsafe_allow_html=True)
response_message = update_formatting(response_message)
chat_box.empty()
st.markdown(response_message, unsafe_allow_html=True)
st.session_state[chat_history_key].append({"content": f"{response_message}", "role": "assistant"})
return response_message
def print_chat_history_timeline(chat_history_key):
for message in st.session_state[chat_history_key]:
role = message["role"]
if role == "user":
with st.chat_message("user", avatar="🧑‍🚀"):
question = message["content"]
# check if content contains arabic characters
question = update_formatting(question)
st.markdown(f"{question}", unsafe_allow_html=True)
elif role == "assistant":
with st.chat_message("assistant", avatar="🤖"):
content = message["content"]
# check if content contains arabic characters
content = update_formatting(content)
st.markdown(content, unsafe_allow_html=True)
# -- helpers --
def page_config():
# page formatting for arabic models
st.set_page_config(layout="wide", page_title="نماذج اللغة للذكاء الأصطناعي بالعربي", page_icon="🦙")
# set entire page to rtl (for arabic support)
st.markdown("<style>body {direction: rtl;}</style>", unsafe_allow_html=True)
st.sidebar.title("تحدث مع اللاما 🦙")
st.sidebar.subheader("نماذج اللغة للذكاء الأصطناعي بالعربي")
footer = """<div style="position: fixed; left: 0; bottom: 0; background-color: #004d4d; color: white; padding: 10px 20px; border-top-right-radius: 10px; box-shadow: 2px 2px 5px grey; font-size: 10px; opacity: 0.5;">
Source code <a href="https://github.com/iamaziz/st_ollama" style="color: white; text-decoration: none; font-weight: bold;">here</a>
<br>
<span style="font-size: 10px;">2023 By Aziz Alto</span>
</div>"""
st.sidebar.markdown(footer, unsafe_allow_html=True)
st.markdown("<style>body {direction: rtl;}</style>", unsafe_allow_html=True)
def select_model():
# model names
llm_names = [model["name"] for model in OLLAMA_MODELS]
llm_name = st.sidebar.selectbox(f"Choose Agent ({len(llm_names)} available)", [""] + llm_names)
# write model details as dict
if llm_name:
llm_details = [model for model in OLLAMA_MODELS if model["name"] == llm_name][0]
with st.expander("Model Details", expanded=False):
st.write(llm_details)
return llm_name
def assert_models_installed():
if len(OLLAMA_MODELS) < 1:
st.sidebar.warning("No models found. Please install at least one model e.g. `ollama run llama2`")
st.stop()
def save_conversation(llm_name, conversation_key):
OUTPUT_DIR = "llm_conversations"
OUTPUT_DIR = os.path.join(os.getcwd(), OUTPUT_DIR)
timestamp = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
filename = f"{OUTPUT_DIR}/{timestamp}_{llm_name.replace(':', '-')}"
if not os.path.exists(OUTPUT_DIR):
os.makedirs(OUTPUT_DIR)
if st.session_state[conversation_key]:
if st.sidebar.button("احفظ المحادثة"):
with open(f"{filename}.json", "w", encoding='utf-8') as f:
json.dump(st.session_state[conversation_key], f, indent=4, ensure_ascii=False)
st.success(f"Conversation saved to {filename}.json")
def is_arabic(text):
"""Check if text contains arabic characters"""
for character in text:
if '\u0600' <= character <= '\u06FF' or '\u0750' <= character <= '\u077F' or '\uFB50' <= character <= '\uFDFF' or '\uFE70' <= character <= '\uFEFF':
return True
return False
if __name__ == "__main__":
page_config()
llm_name = select_model()
assert_models_installed()
if not llm_name: st.stop()
conversation_key = f"model_{llm_name}"
ask_question_message = f"أسألني عن أي شيء"
prompt = st.chat_input(ask_question_message)
st_ollama(llm_name, prompt, conversation_key)
if st.session_state[conversation_key]:
clear_conversation = st.sidebar.button("امسح المحادثة")
if clear_conversation:
st.session_state[conversation_key] = []
st.rerun()
# save conversation to file
save_conversation(llm_name, conversation_key)
@iamaziz
Copy link
Author

iamaziz commented Feb 28, 2024

Example

arabic-ollachat-ui.mov

@jaber1988
Copy link

where to find this arabic_nous hermes2 ?? please tell us ?

@iamaziz
Copy link
Author

iamaziz commented Apr 17, 2024

I downloaded it from Ollama, and used the system prompt in this notebook.

However, I would recommend using command-r or command-r-plus, they have the best out-of-the-box support for Arabic language I've seen (for open LLMs).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment