rlhf-interface / app.py
Tristan Thrush
synched model memory with conversation, and make sure it is wiped for next hit
90b6f98
raw
history blame
11.3 kB
# Basic example for doing model-in-the-loop dynamic adversarial data collection
# using Gradio Blocks.
import json
import os
import threading
import uuid
from pathlib import Path
from urllib.parse import parse_qs
import gradio as gr
from dotenv import load_dotenv
from huggingface_hub import Repository
from langchain import ConversationChain
from langchain.chains.conversation.memory import ConversationBufferMemory
from langchain.llms import HuggingFaceHub
from langchain.prompts import load_prompt
from utils import force_git_push
# These variables are for storing the mturk HITs in a Hugging Face dataset.
if Path(".env").is_file():
load_dotenv(".env")
DATASET_REPO_URL = os.getenv("DATASET_REPO_URL")
FORCE_PUSH = os.getenv("FORCE_PUSH")
HF_TOKEN = os.getenv("HF_TOKEN")
PROMPT_TEMPLATES = Path("prompt_templates")
# Set env variable for langchain to communicate with Hugging Face Hub
os.environ["HUGGINGFACEHUB_API_TOKEN"] = HF_TOKEN
DATA_FILENAME = "data.jsonl"
DATA_FILE = os.path.join("data", DATA_FILENAME)
repo = Repository(
local_dir="data", clone_from=DATASET_REPO_URL, use_auth_token=HF_TOKEN
)
TOTAL_CNT = 3 # How many user inputs per HIT
# This function pushes the HIT data written in data.jsonl to our Hugging Face
# dataset every minute. Adjust the frequency to suit your needs.
PUSH_FREQUENCY = 60
def asynchronous_push(f_stop):
if repo.is_repo_clean():
print("Repo currently clean. Ignoring push_to_hub")
else:
repo.git_add(auto_lfs_track=True)
repo.git_commit("Auto commit by space")
if FORCE_PUSH == "yes":
force_git_push(repo)
else:
repo.git_push()
if not f_stop.is_set():
# call again in 60 seconds
threading.Timer(PUSH_FREQUENCY, asynchronous_push, [f_stop]).start()
f_stop = threading.Event()
asynchronous_push(f_stop)
# Now let's run the app!
prompt = load_prompt(PROMPT_TEMPLATES / "openai_chatgpt.json")
chatbot_1 = ConversationChain(
llm=HuggingFaceHub(
repo_id="google/flan-t5-xl",
model_kwargs={"temperature": 1}
),
prompt=prompt,
verbose=False,
memory=ConversationBufferMemory(ai_prefix="Assistant"),
)
chatbot_2 = ConversationChain(
llm=HuggingFaceHub(
repo_id="bigscience/bloom",
model_kwargs={"temperature": 0.7}
),
prompt=prompt,
verbose=False,
memory=ConversationBufferMemory(ai_prefix="Assistant"),
)
chatbot_3 = ConversationChain(
llm=HuggingFaceHub(
repo_id="bigscience/T0_3B",
model_kwargs={"temperature": 1}
),
prompt=prompt,
verbose=False,
memory=ConversationBufferMemory(ai_prefix="Assistant"),
)
chatbot_4 = ConversationChain(
llm=HuggingFaceHub(
repo_id="EleutherAI/gpt-j-6B",
model_kwargs={"temperature": 1}
),
prompt=prompt,
verbose=False,
memory=ConversationBufferMemory(ai_prefix="Assistant"),
)
model_id2model = {
"google/flan-t5-xl": chatbot_1,
"bigscience/bloom": chatbot_2,
"bigscience/T0_3B": chatbot_3,
"EleutherAI/gpt-j-6B": chatbot_4
}
demo = gr.Blocks()
with demo:
dummy = gr.Textbox(visible=False) # dummy for passing assignmentId
# We keep track of state as a JSON
state_dict = {
"conversation_id": str(uuid.uuid4()),
"assignmentId": "",
"cnt": 0, "data": [],
"past_user_inputs": [],
"generated_responses": [],
"response_1": "",
"response_2": "",
"response_3": "",
"response_4": "",
}
state = gr.JSON(state_dict, visible=False)
gr.Markdown("# RLHF Interface")
gr.Markdown("Choose the best model output")
state_display = gr.Markdown(f"Your messages: 0/{TOTAL_CNT}")
# Generate model prediction
def _predict(txt, state):
# TODO: parallelize this!
response_1 = chatbot_1.predict(input=txt)
response_2 = chatbot_2.predict(input=txt)
response_3 = chatbot_3.predict(input=txt)
response_4 = chatbot_4.predict(input=txt)
response2model_id = {}
response2model_id[response_1] = chatbot_1.llm.repo_id
response2model_id[response_2] = chatbot_2.llm.repo_id
response2model_id[response_3] = chatbot_3.llm.repo_id
response2model_id[response_4] = chatbot_4.llm.repo_id
state["cnt"] += 1
new_state_md = f"Inputs remaining in HIT: {state['cnt']}/{TOTAL_CNT}"
state["data"].append({"cnt": state["cnt"], "text": txt, "response_1": response_1, "response_2": response_2, "response_3": response_3, "response_4": response_4,"response2model_id": response2model_id})
state["past_user_inputs"].append(txt)
past_conversation_string = "<br />".join(["<br />".join(["πŸ˜ƒ: " + user_input, "πŸ€–: " + model_response]) for user_input, model_response in zip(state["past_user_inputs"], state["generated_responses"] + [""])])
return gr.update(visible=False), gr.update(visible=True), gr.update(visible=True, choices=[response_1, response_2, response_3, response_4], interactive=True, value=response_1), gr.update(value=past_conversation_string), state, gr.update(visible=False), gr.update(visible=False), gr.update(visible=False), new_state_md, dummy
def _select_response(selected_response, state, dummy):
done = state["cnt"] == TOTAL_CNT
state["generated_responses"].append(selected_response)
state["data"][-1]["selected_response"] = selected_response
state["data"][-1]["selected_model"] = state["data"][-1]["response2model_id"][selected_response]
if state["cnt"] == TOTAL_CNT:
# Write the HIT data to our local dataset because the worker has
# submitted everything now.
with open(DATA_FILE, "a") as jsonlfile:
json_data_with_assignment_id =\
[json.dumps(dict({"assignmentId": state["assignmentId"], "conversation_id": state["conversation_id"]}, **datum)) for datum in state["data"]]
jsonlfile.write("\n".join(json_data_with_assignment_id) + "\n")
toggle_example_submit = gr.update(visible=not done)
past_conversation_string = "<br />".join(["<br />".join(["πŸ˜ƒ: " + user_input, "πŸ€–: " + model_response]) for user_input, model_response in zip(state["past_user_inputs"], state["generated_responses"])])
query = parse_qs(dummy[1:])
if "assignmentId" in query and query["assignmentId"][0] != "ASSIGNMENT_ID_NOT_AVAILABLE":
# It seems that someone is using this app on mturk. We need to
# store the assignmentId in the state before submit_hit_button
# is clicked. We can do this here in _predict. We need to save the
# assignmentId so that the turker can get credit for their HIT.
state["assignmentId"] = query["assignmentId"][0]
toggle_final_submit = gr.update(visible=done)
toggle_final_submit_preview = gr.update(visible=False)
else:
toggle_final_submit_preview = gr.update(visible=done)
toggle_final_submit = gr.update(visible=False)
if done:
# Wipe the memory completely because we will be starting a new hit soon.
chatbot_1.memory = ConversationBufferMemory(ai_prefix="Assistant")
chatbot_2.memory = ConversationBufferMemory(ai_prefix="Assistant")
chatbot_3.memory = ConversationBufferMemory(ai_prefix="Assistant")
chatbot_4.memory = ConversationBufferMemory(ai_prefix="Assistant")
else:
# Sync all of the model's memories with the conversation path that
# was actually taken.
chatbot_1.memory = model_id2model[state["data"][-1]["response2model_id"][selected_response]].memory
chatbot_2.memory = model_id2model[state["data"][-1]["response2model_id"][selected_response]].memory
chatbot_3.memory = model_id2model[state["data"][-1]["response2model_id"][selected_response]].memory
chatbot_4.memory = model_id2model[state["data"][-1]["response2model_id"][selected_response]].memory
text_input = gr.update(visible=False) if done else gr.update(visible=True)
return gr.update(visible=False), gr.update(visible=True), text_input, gr.update(visible=False), state, gr.update(value=past_conversation_string), toggle_example_submit, toggle_final_submit, toggle_final_submit_preview,
# Input fields
past_conversation = gr.Markdown()
text_input = gr.Textbox(placeholder="Enter a statement", show_label=False)
select_response = gr.Radio(choices=[None, None], visible=False, label="Choose the best response")
select_response_button = gr.Button("Select Response", visible=False)
with gr.Column() as example_submit:
submit_ex_button = gr.Button("Submit")
with gr.Column(visible=False) as final_submit:
submit_hit_button = gr.Button("Submit HIT")
with gr.Column(visible=False) as final_submit_preview:
submit_hit_button_preview = gr.Button("Submit Work (preview mode; no mturk HIT credit, but your examples will still be stored)")
# Button event handlers
get_window_location_search_js = """
function(text_input, label_input, state, dummy) {
return [text_input, label_input, state, window.location.search];
}
"""
select_response_button.click(
_select_response,
inputs=[select_response, state, dummy],
outputs=[select_response, example_submit, text_input, select_response_button, state, past_conversation, example_submit, final_submit, final_submit_preview],
_js=get_window_location_search_js,
)
submit_ex_button.click(
_predict,
inputs=[text_input, state],
outputs=[text_input, select_response_button, select_response, past_conversation, state, example_submit, final_submit, final_submit_preview, state_display, dummy],
_js=get_window_location_search_js,
)
post_hit_js = """
function(state) {
// If there is an assignmentId, then the submitter is on mturk
// and has accepted the HIT. So, we need to submit their HIT.
const form = document.createElement('form');
form.action = 'https://workersandbox.mturk.com/mturk/externalSubmit';
form.method = 'post';
for (const key in state) {
const hiddenField = document.createElement('input');
hiddenField.type = 'hidden';
hiddenField.name = key;
hiddenField.value = state[key];
form.appendChild(hiddenField);
};
document.body.appendChild(form);
form.submit();
return state;
}
"""
submit_hit_button.click(
lambda state: state,
inputs=[state],
outputs=[state],
_js=post_hit_js,
)
refresh_app_js = """
function(state) {
// The following line here loads the app again so the user can
// enter in another preview-mode "HIT".
window.location.href = window.location.href;
return state;
}
"""
submit_hit_button_preview.click(
lambda state: state,
inputs=[state],
outputs=[state],
_js=refresh_app_js,
)
demo.launch()