grahamwhiteuk's picture
fix: deployment
026d799 verified
raw
history blame
12 kB
import json
import os
import gradio as gr
from dotenv import load_dotenv
from gradio_modal import Modal
from logger import logger
from model import generate_text, get_prompt
from utils import (
get_messages,
get_result_description,
load_command_line_args,
to_snake_case,
to_title_case,
)
load_command_line_args()
load_dotenv()
catalog = {}
with open("catalog.json") as f:
logger.debug("Loading catalog from json.")
catalog = json.load(f)
def update_selected_test_case(button_name, state: gr.State, event: gr.EventData):
target_sub_catalog_name, target_test_case_name = event.target.elem_id.split("---")
state["selected_sub_catalog"] = target_sub_catalog_name
state["selected_criteria_name"] = target_test_case_name
state["selected_test_case"] = [
t
for sub_catalog_name, sub_catalog in catalog.items()
for t in sub_catalog
if t["name"] == to_snake_case(button_name) and to_snake_case(sub_catalog_name) == target_sub_catalog_name
][0]
return state
def on_test_case_click(state: gr.State):
selected_sub_catalog = state["selected_sub_catalog"]
selected_criteria_name = state["selected_criteria_name"]
selected_test_case = state["selected_test_case"]
logger.debug(f'Changing to test case "{selected_criteria_name}" from catalog "{selected_sub_catalog}".')
is_context_iditable = selected_criteria_name == "context_relevance"
is_user_message_editable = selected_sub_catalog == "harmful_content_in_user_prompt"
is_assistant_message_editable = (
selected_sub_catalog == "harmful_content_in_assistant_response"
or selected_criteria_name == "groundedness"
or selected_criteria_name == "answer_relevance"
)
return {
test_case_name: f'<h2>{to_title_case(selected_test_case["name"])}</h2>',
criteria: selected_test_case["criteria"],
context: (
gr.update(value=selected_test_case["context"], interactive=True, visible=True, elem_classes=["input-box"])
if is_context_iditable
else gr.update(
visible=selected_test_case["context"] is not None,
value=selected_test_case["context"],
interactive=False,
elem_classes=["read-only", "input-box"],
)
),
user_message: (
gr.update(
value=selected_test_case["user_message"], visible=True, interactive=True, elem_classes=["input-box"]
)
if is_user_message_editable
else gr.update(
value=selected_test_case["user_message"], interactive=False, elem_classes=["read-only", "input-box"]
)
),
assistant_message: (
gr.update(
value=selected_test_case["assistant_message"],
visible=True,
interactive=True,
elem_classes=["input-box"],
)
if is_assistant_message_editable
else gr.update(
visible=selected_test_case["assistant_message"] is not None,
value=selected_test_case["assistant_message"],
interactive=False,
elem_classes=["read-only", "input-box"],
)
),
result_text: gr.update(visible=False, value=""),
}
def change_button_color(event: gr.EventData):
return [
(
gr.update(elem_classes=["catalog-button", "selected"])
if v.elem_id == event.target.elem_id
else gr.update(elem_classes=["catalog-button"])
)
for c in catalog_buttons.values()
for v in c.values()
]
def on_submit(criteria, context, user_message, assistant_message, state):
criteria_name = state["selected_criteria_name"]
test_case = {
"name": criteria_name,
"criteria": criteria,
"context": context,
"user_message": user_message,
"assistant_message": assistant_message,
}
messages = get_messages(test_case=test_case, sub_catalog_name=state["selected_sub_catalog"])
logger.debug(
f"Starting evaluation for subcatelog {state['selected_sub_catalog']} and criteria name {state['selected_criteria_name']}"
)
result_label = generate_text(messages=messages, criteria_name=criteria_name)["assessment"] # Yes or No
html_str = f"<p>{get_result_description(state['selected_sub_catalog'], state['selected_criteria_name'])} <strong>{result_label}</strong></p>"
# html_str = f"{get_result_description(state['selected_sub_catalog'], state['selected_criteria_name'])} {result_label}"
return gr.update(value=html_str)
def on_show_prompt_click(criteria, context, user_message, assistant_message, state):
criteria_name = state["selected_criteria_name"]
test_case = {
"name": criteria_name,
"criteria": criteria,
"context": context,
"user_message": user_message,
"assistant_message": assistant_message,
}
messages = get_messages(test_case=test_case, sub_catalog_name=state["selected_sub_catalog"])
prompt = get_prompt(messages, criteria_name)
print(prompt)
prompt = prompt.replace("<", "&lt;").replace(">", "&gt;").replace("\\n", "<br>")
return gr.Markdown(prompt)
ibm_blue = gr.themes.Color(
name="ibm-blue",
c50="#eff6ff",
c100="#dbeafe",
c200="#bfdbfe",
c300="#93c5fd",
c400="#60a5fa",
c500="#0F62FE",
c600="#2563eb",
c700="#1d4ed8",
c800="#1e40af",
c900="#1e3a8a",
c950="#1d3660",
)
head_style = """
<style>
@media (min-width: 1536px)
{
.gradio-container {
max-width: 1124px !important;
}
}
@media (min-width: 1024px)
{
.gradio-container {
max-width: 1124px !important;
}
}
</style>
"""
with gr.Blocks(
title="Granite Guardian",
theme=gr.themes.Soft(
primary_hue=ibm_blue,
font=[gr.themes.GoogleFont("IBM Plex Sans"), gr.themes.GoogleFont("Source Sans 3")],
),
head=head_style,
fill_width=False,
css=os.path.join(os.path.dirname(os.path.abspath(__file__)), "./styles.css"),
) as demo:
state = gr.State(
value={"selected_sub_catalog": "harmful_content_in_user_prompt", "selected_criteria_name": "general_harm"}
)
starting_test_case = [
t
for sub_catalog_name, sub_catalog in catalog.items()
for t in sub_catalog
if t["name"] == state.value["selected_criteria_name"]
and sub_catalog_name == state.value["selected_sub_catalog"]
][0]
with gr.Row(elem_classes="header-row"):
with gr.Column(scale=4):
gr.HTML("<h2>IBM Granite Guardian 3.0</h2>", elem_classes="title")
gr.HTML(
elem_classes="system-description",
value="<p>Granite Guardian models are specialized language models in the Granite family that can detect harms and risks in generative AI systems. They can be used with any large language model to make interactions with generative AI systems safe. Select an example in the left panel to see how the Granite Guardian model evaluates harms and risks in user prompts, assistant responses, and for hallucinations in retrieval-augmented generation. In this demo, we use granite-guardian-3.0-8b.</p>",
)
with gr.Row(elem_classes="column-gap"):
with gr.Column(scale=0, elem_classes="no-gap"):
title_display_left = gr.HTML("<h2>Harms & Risks</h2>", elem_classes=["subtitle", "subtitle-harms"])
accordions = []
catalog_buttons: dict[str, dict[str, gr.Button]] = {}
for i, (sub_catalog_name, sub_catalog) in enumerate(catalog.items()):
with gr.Accordion(
to_title_case(sub_catalog_name), open=(i == 0), elem_classes="accordion"
) as accordion:
for test_case in sub_catalog:
elem_classes = ["catalog-button"]
elem_id = f"{sub_catalog_name}---{test_case['name']}"
if starting_test_case == test_case:
elem_classes.append("selected")
if sub_catalog_name not in catalog_buttons:
catalog_buttons[sub_catalog_name] = {}
catalog_buttons[sub_catalog_name][test_case["name"]] = gr.Button(
to_title_case(test_case["name"]),
elem_classes=elem_classes,
variant="secondary",
size="sm",
elem_id=elem_id,
)
accordions.append(accordion)
with gr.Column(visible=True, scale=1) as test_case_content:
with gr.Row(elem_classes="no-stretch"):
test_case_name = gr.HTML(
f'<h2>{to_title_case(starting_test_case["name"])}</h2>', elem_classes="subtitle"
)
show_propt_button = gr.Button("Show prompt", size="sm", scale=0, min_width=110)
criteria = gr.Textbox(
label="Evaluation Criteria",
lines=3,
interactive=False,
value=starting_test_case["criteria"],
elem_classes=["read-only", "input-box", "margin-bottom"],
)
gr.HTML(elem_classes=["block", "content-gap"])
context = gr.Textbox(
label="Context",
lines=3,
interactive=True,
value=starting_test_case["context"],
visible=False,
elem_classes=["input-box"],
)
user_message = gr.Textbox(
label="User Prompt",
lines=3,
interactive=True,
value=starting_test_case["user_message"],
elem_classes=["input-box"],
)
assistant_message = gr.Textbox(
label="Assistant Response",
lines=3,
interactive=True,
visible=False,
value=starting_test_case["assistant_message"],
elem_classes=["input-box"],
)
submit_button = gr.Button(
"Evaluate",
variant="primary",
icon=os.path.join(os.path.dirname(os.path.abspath(__file__)), "send-white.png"),
elem_classes="submit-button",
)
# result_text = gr.HTML(label='Result', elem_classes=['result-text', 'read-only', 'input-box'], visible=False, value='')
result_text = gr.HTML(
label="Result", elem_classes=["result-root"], show_label=True, visible=False, value=""
)
with Modal(visible=False, elem_classes="modal") as modal:
prompt = gr.Markdown("")
# events
show_propt_button.click(
on_show_prompt_click, inputs=[criteria, context, user_message, assistant_message, state], outputs=prompt
).then(lambda: gr.update(visible=True), None, modal)
submit_button.click(lambda: gr.update(visible=True, value=""), None, result_text).then(
on_submit,
inputs=[criteria, context, user_message, assistant_message, state],
outputs=[result_text],
scroll_to_output=True,
)
for button in [
t for sub_catalog_name, sub_catalog_buttons in catalog_buttons.items() for t in sub_catalog_buttons.values()
]:
button.click(
change_button_color, inputs=None, outputs=[v for c in catalog_buttons.values() for v in c.values()]
).then(update_selected_test_case, inputs=[button, state], outputs=[state]).then(
on_test_case_click,
inputs=state,
outputs={test_case_name, criteria, context, user_message, assistant_message, result_text},
)
demo.launch(server_name="0.0.0.0")