import datetime
import random
import os
import re
from io import StringIO
import gradio as gr
import pandas as pd
from huggingface_hub import Repository, upload_file
from text_generation import Client
from share_btn import (community_icon_html, loading_icon_html, share_btn_css,
share_js)
HF_TOKEN = os.environ.get("HF_TOKEN", None)
API_TOKEN = os.environ.get("API_TOKEN", HF_TOKEN)
DIALOGUES_DATASET = "ArmelR/gradio_playground_dialogues"
API_URL_G = "https://api-inference.huggingface.co/models/ArmelR/starcoder-gradio-v0"
API_URL_S = "https://api-inference.huggingface.co/models/HuggingFaceH4/starcoderbase-finetuned-oasst1"
API_URL_B = "https://api-inference.huggingface.co/models/HuggingFaceH4/starchat-beta"
model2endpoint = {
"starChat-alpha": API_URL_S,
"starCoder-gradio": API_URL_G,
"starChat-beta": API_URL_B
}
model_names = list(model2endpoint.keys())
with open("./HHH_prompt_short.txt", "r") as f:
HHH_PROMPT = f.read() + "\n\n"
with open("./TA_prompt_v0.txt", "r") as f:
TA_PROMPT = f.read()
NO_PROMPT = ""
def randomize_seed_generator():
seed = random.randint(0, 1000000)
return seed
def save_inputs_and_outputs(now, inputs, outputs, generate_kwargs, model):
buffer = StringIO()
timestamp = datetime.datetime.now().strftime("%Y-%m-%dT%H:%M:%S.%f")
file_name = f"prompts_{timestamp}.jsonl"
data = {"model": model, "inputs": inputs, "outputs": outputs, "generate_kwargs": generate_kwargs}
pd.DataFrame([data]).to_json(buffer, orient="records", lines=True)
# Push to Hub
upload_file(
path_in_repo=f"{now.date()}/{now.hour}/{file_name}",
path_or_fileobj=buffer.getvalue().encode(),
repo_id=DIALOGUES_DATASET,
token=HF_TOKEN,
repo_type="dataset",
)
# Clean and rerun
buffer.close()
def get_total_inputs(inputs, chatbot, preprompt, user_name, assistant_name, sep):
past = []
for data in chatbot:
user_data, model_data = data
if not user_data.startswith(user_name):
user_data = user_name + user_data
if not model_data.startswith(sep + assistant_name):
model_data = sep + assistant_name + model_data
past.append(user_data + model_data.rstrip() + sep)
if not inputs.startswith(user_name):
inputs = user_name + inputs
total_inputs = preprompt + "".join(past) + inputs + sep + assistant_name.rstrip()
return total_inputs
def wrap_html_code(text):
pattern = r"<.*?>"
matches = re.findall(pattern, text)
if len(matches) > 0:
return f"```{text}```"
else:
return text
def has_no_history(chatbot, history):
return not chatbot and not history
def get_inference_prompt(messages, model_name):
if model_name == "starChat-beta" :
prompt = "<|system|>\n<|endoftext|>\n"
for message in messages :
if message["role"] == "user" :
prompt += f"<|user|>\n{message['content']}<|endoftext|>\n<|assistant|>"
else : # message["role"] == "assistant"
prompt += f"\n{message['content']}<|endoftext|>\n"
elif model_name == "starChat-alpha" :
prompt = "<|system|>\n<|end|>\n"
for message in messages :
if message["role"] == "user" :
prompt += f"<|user|>\n{message['content']}<|end|>\n<|assistant|>"
else : # message["role"] == "assistant"
prompt += f"\n{message['content']}<|end|>\n"
else : # starCoder-gradio
prompt = ""
for message in messages :
if message["role"] == "user" :
prompt += f"Question: {message['content']}\n\nAnswer:"
else : # message["role"] == "assistant"
prompt += f" {message['content']}\n\n"
return prompt
def generate(
RETRY_FLAG,
model_name,
system_message,
user_message,
chatbot,
history,
temperature,
top_k,
top_p,
max_new_tokens,
repetition_penalty,
do_save=True,
):
client = Client(
model2endpoint[model_name],
headers={"Authorization": f"Bearer {API_TOKEN}"},
timeout=60,
)
# Don't return meaningless message when the input is empty
if not user_message:
print("Empty input")
if not RETRY_FLAG:
history.append(user_message)
seed = 42
else:
seed = randomize_seed_generator()
past_messages = []
for data in chatbot:
user_data, model_data = data
past_messages.extend(
[{"role": "user", "content": user_data}, {"role": "assistant", "content": model_data.rstrip()}]
)
if len(past_messages) < 1:
prompt = get_inference_prompt(messages=[{"role": "user", "content": user_message}], model_name=model_name)
else:
prompt = get_inference_prompt(messages=past_messages + [{"role": "user", "content": user_message}], model_name=model_name)
generate_kwargs = {
"temperature": temperature,
"top_k": top_k,
"top_p": top_p,
"max_new_tokens": max_new_tokens,
}
temperature = float(temperature)
if temperature < 1e-2:
temperature = 1e-2
top_p = float(top_p)
generate_kwargs = dict(
temperature=temperature,
max_new_tokens=max_new_tokens,
top_p=top_p,
repetition_penalty=repetition_penalty,
do_sample=True,
truncate=4096,
seed=seed,
stop_sequences=["<|end|>", "Question:"],
)
stream = client.generate_stream(
prompt,
**generate_kwargs,
)
output = ""
for idx, response in enumerate(stream):
if response.token.special:
continue
output += response.token.text
if idx == 0:
history.append(" " + output)
else:
history[-1] = output
chat = [
(wrap_html_code(history[i].strip()), wrap_html_code(history[i + 1].strip()))
for i in range(0, len(history) - 1, 2)
]
# chat = [(history[i].strip(), history[i + 1].strip()) for i in range(0, len(history) - 1, 2)]
yield chat, history, user_message, ""
if HF_TOKEN and do_save:
try:
now = datetime.datetime.now()
current_time = now.strftime("%Y-%m-%d %H:%M:%S")
print(f"[{current_time}] Pushing prompt and completion to the Hub")
save_inputs_and_outputs(now, prompt, output, generate_kwargs, model_name)
except Exception as e:
print(e)
return chat, history, user_message, ""
examples = [
"""How to upload dataframe with content of file upload? I did not find any documentation of this.
I want to upload a CSV file, display the data in an interactive table and then us the table conents to create a plot. But I am stuck with gluing the components together.
import gradio as gr
default_csv = "Phase,Activity,Start date,End date\n\"Mapping the Field\",\"Literature review\",2024-01-01,2024-01-31"
def process_csv_text(text):
df = pd.read_csv(StringIO(text), parse_dates=["Start date", "End date"])
return df
with gr.Blocks() as demo:
upload_button = gr.UploadButton(label="Upload Timetable", file_types = ['.csv'], live=True, file_count = "single")
table = gr.Dataframe(headers=["Phase", "Activity", "Start date", "End date"], col_count=4, default=process_csv_text(default_csv))
image = gr.Plot()
upload_button.click(fn=process_csv_text, inputs=upload_button, outputs=table, api_name="upload_csv")
demo.launch()
""",
"""Hi, I want to remove the "clear" button in Gradio Interface. Is there a way to do such a thing?""",
"""Why the botton can not apply css? But textbox works great.
import gradio as gr
import numpy as np
css = \"""
#warning {background-color: #FFCCCB}
.feedback textarea {font-size: 64px !important}
\"""
with gr.Blocks(css=css) as demo:
box1 = gr.Button(value="Good Job", elem_classes="feedback")
box2 = gr.Textbox(value="Failure", elem_id="warning", elem_classes="feedback")
if __name__ == "__main__":
demo.launch(inbrowser=True) """,
"""
I'm very new to both huggingface and Gradio, so forgive me if this is a trivial issue, and I'm just a fool.
I'm using a Blocks object to implement a very basic chatbot, and I'd like to essentially clear the input textbox after the user presses enter or presses the submit button.
with gr.Blocks() as test:
outp = gr.Chatbot(label="Reply", )
inp = gr.Textbox(label="Chat with AI")
inp.submit(chatbot, [outp, inp], outp)
inp.update(lambda: None, inp)
btn = gr.Button("Submit")
btn.click(fn=chatbot, inputs=inp, outputs=outp)
My code looks like this currently.
I simply wish to clear the input textbox after the input is submitted.
""",
"""
I am trying to force Gradio to show a white background colour ALL the time in all browsers. My code below works to force white colour on my desktop (using Firefox) but on mobile it's still showing the typical default Gradio black background. How to change this behaviour to permanently show a white background in all devices? Thank you
```python
demo = gr.Interface(
lambda x:x+x,
inputs=gr.Textbox(label='Test'),
outputs=gr.Textbox(label='test2'),
css=".gradio-container {background-color: white} "
).launch(share=False)
```
""",
"""
I am using gradio. I'm creating some tabs and when i click a button i want to use the name of the tabs as an input for a function. How does it work? Thank you very much!
My way of creating tabs would be for example this, with blocks:
```python
with gr.Tabs() as tabs:
with gr.TabItem("test1"):
bt_test1 = gr.Button('test1')
with gr.TabItem("test2"):
bt_test2 = gr.Button('test2')
```
""",
"""
Can I change the layout when I click a button in gradio?
For example, I would like to add a row or change some other layout of the frame when i click a button. how can i do this operation.
"""
]
def clear_chat():
return [], []
def delete_last_turn(chat, history):
if chat and history:
chat.pop(-1)
history.pop(-1)
history.pop(-1)
return chat, history
def process_example(args):
for [x, y] in generate(args):
pass
return [x, y]
# Regenerate response
def retry_last_answer(
selected_model,
system_message,
user_message,
chat,
history,
temperature,
top_k,
top_p,
max_new_tokens,
repetition_penalty,
do_save,
):
if chat and history:
# Removing the previous conversation from chat
chat.pop(-1)
# Removing bot response from the history
history.pop(-1)
# Setting up a flag to capture a retry
RETRY_FLAG = True
# Getting last message from user
user_message = history[-1]
yield from generate(
RETRY_FLAG,
selected_model,
system_message,
user_message,
chat,
history,
temperature,
top_k,
top_p,
max_new_tokens,
repetition_penalty,
do_save,
)
title = """Chat with Gradio 💫➕
"""
custom_css = """
#banner-image {
display: block;
margin-left: auto;
margin-right: auto;
}
#chat-message {
font-size: 14px;
min-height: 300px;
}
"""
with gr.Blocks(analytics_enabled=False, css=custom_css) as demo:
#gr.HTML(title)
with gr.Row():
#with gr.Column():
#gr.Image("gradio.png", elem_id="banner-image", show_label=False)
#with gr.Column():
# gr.Markdown(
# """
# 💻 This demo showcases a series of **[StarChat](https://huggingface.co/models?search=huggingfaceh4/starchat)** language models, which are fine-tuned versions of the StarCoder family to act as helpful coding assistants. The base model has 16B parameters and was pretrained on one trillion tokens sourced from 80+ programming languages, GitHub issues, Git commits, and Jupyter notebooks (all permissively licensed).
# 📝 For more details, check out our [blog post](https://huggingface.co/blog/starchat-alpha).
# ⚠️ **Intended Use**: this app and its [supporting models](https://huggingface.co/models?search=huggingfaceh4/starchat) are provided as educational tools to explain large language model fine-tuning; not to serve as replacement for human expertise.
# ⚠️ **Known Failure Modes**: the alpha and beta version of **StarChat** have not been aligned to human preferences with techniques like RLHF, so they can produce problematic outputs (especially when prompted to do so). Since the base model was pretrained on a large corpus of code, it may produce code snippets that are syntactically valid but semantically incorrect. For example, it may produce code that does not compile or that produces incorrect results. It may also produce code that is vulnerable to security exploits. We have observed the model also has a tendency to produce false URLs which should be carefully inspected before clicking. For more details on the model's limitations in terms of factuality and biases, see the [model card](https://huggingface.co/HuggingFaceH4/starchat-alpha#bias-risks-and-limitations).
# ⚠️ **Data Collection**: by default, we are collecting the prompts entered in this app to further improve and evaluate the models. Do **NOT** share any personal or sensitive information while using the app! You can opt out of this data collection by removing the checkbox below.
# """
# )
with gr.Column():
gr.Markdown("""
""")
with gr.Row():
selected_model = gr.Radio(choices=model_names, value=model_names[1], label="Select a model")
with gr.Accordion(label="System Prompt", open=False, elem_id="parameters-accordion"):
system_message = gr.Textbox(
elem_id="system-message",
placeholder="Below is a conversation between a human user and a helpful AI coding assistant.",
show_label=False,
)
with gr.Row():
with gr.Box():
output = gr.Markdown()
chatbot = gr.Chatbot(elem_id="chat-message", label="Chat")
with gr.Row():
with gr.Column(scale=3):
user_message = gr.Textbox(placeholder="Enter your message here", show_label=False, elem_id="q-input")
with gr.Row():
send_button = gr.Button("Send", elem_id="send-btn", visible=True)
regenerate_button = gr.Button("Regenerate", elem_id="retry-btn", visible=True)
delete_turn_button = gr.Button("Delete last turn", elem_id="delete-btn", visible=True)
clear_chat_button = gr.Button("Clear chat", elem_id="clear-btn", visible=True)
with gr.Accordion(label="Parameters", open=False, elem_id="parameters-accordion"):
temperature = gr.Slider(
label="Temperature",
value=0.2,
minimum=0.0,
maximum=1.0,
step=0.1,
interactive=True,
info="Higher values produce more diverse outputs",
)
top_k = gr.Slider(
label="Top-k",
value=50,
minimum=0.0,
maximum=100,
step=1,
interactive=True,
info="Sample from a shortlist of top-k tokens",
)
top_p = gr.Slider(
label="Top-p (nucleus sampling)",
value=0.95,
minimum=0.0,
maximum=1,
step=0.05,
interactive=True,
info="Higher values sample more low-probability tokens",
)
max_new_tokens = gr.Slider(
label="Max new tokens",
value=512,
minimum=0,
maximum=1024,
step=4,
interactive=True,
info="The maximum numbers of new tokens",
)
repetition_penalty = gr.Slider(
label="Repetition Penalty",
value=1.2,
minimum=0.0,
maximum=10,
step=0.1,
interactive=True,
info="The parameter for repetition penalty. 1.0 means no penalty.",
)
with gr.Row():
do_save = gr.Checkbox(
value=True,
label="Store data",
info="You agree to the storage of your prompt and generated text for research and development purposes:",
)
# with gr.Group(elem_id="share-btn-container"):
# community_icon = gr.HTML(community_icon_html, visible=True)
# loading_icon = gr.HTML(loading_icon_html, visible=True)
# share_button = gr.Button("Share to community", elem_id="share-btn", visible=True)
with gr.Row():
gr.Examples(
examples=examples,
inputs=[user_message],
cache_examples=False,
fn=process_example,
outputs=[output],
)
with gr.Row():
gr.Markdown(
"""Chat-with-Gradio is a 15.5 billion parameter language model based on [BigCode's StarCoderplus model](https://huggingface.co/bigcode/starcoderplus) that has been trained on a wide variety of data sources. It includes the source code and issues from [gradio's Github repository](https://github.com/gradio-app/gradio) and data from [Hugging Face's spaces](https://huggingface.co/spaces). Its training also involves instruction fine-tuning with a processed subset of [OpenAssistant's oasst1 dataset](https://huggingface.co/datasets/HuggingFaceH4/oasst1_en).
Type in the box below and click the button to generate answers to your most pressing questions!
⚠️ **Intended Use**: this app and its [supporting model](https://huggingface.co/bigcode/starcoderplus) are provided as tools to provide assistance when using gradio ; not to serve as replacement for human expertise. For more details on the model's limitations in terms of factuality and biases, see the [model card.](https://huggingface.co/bigcode/starcoderplus#intended-uses--limitations)
⚠️ **Data Collection**: by default, we are collecting the prompts entered in this app to further improve and evaluate the model. Do not share any personal or sensitive information while using the app! You can opt out of this data collection by removing the checkbox below:)""")
history = gr.State([])
RETRY_FLAG = gr.Checkbox(value=False, visible=False)
# To clear out "message" input textbox and use this to regenerate message
last_user_message = gr.State("")
user_message.submit(
generate,
inputs=[
RETRY_FLAG,
selected_model,
system_message,
user_message,
chatbot,
history,
temperature,
top_k,
top_p,
max_new_tokens,
repetition_penalty,
do_save,
],
outputs=[chatbot, history, last_user_message, user_message],
)
send_button.click(
generate,
inputs=[
RETRY_FLAG,
selected_model,
system_message,
user_message,
chatbot,
history,
temperature,
top_k,
top_p,
max_new_tokens,
repetition_penalty,
do_save,
],
outputs=[chatbot, history, last_user_message, user_message],
)
regenerate_button.click(
retry_last_answer,
inputs=[
selected_model,
system_message,
user_message,
chatbot,
history,
temperature,
top_k,
top_p,
max_new_tokens,
repetition_penalty,
do_save,
],
outputs=[chatbot, history, last_user_message, user_message],
)
delete_turn_button.click(delete_last_turn, [chatbot, history], [chatbot, history])
clear_chat_button.click(clear_chat, outputs=[chatbot, history])
selected_model.change(clear_chat, outputs=[chatbot, history])
# share_button.click(None, [], [], _js=share_js)
demo.queue(concurrency_count=16).launch(debug=True)