compl-ai-board / app.py
pavol-bielik's picture
add principles and technical requirements mapping
b615923
import functools
from pathlib import Path
import gradio as gr
import pandas as pd
from huggingface_hub import snapshot_download
from src.display.about import (
CITATION_BUTTON_LABEL,
CITATION_BUTTON_TEXT,
EVALUATION_QUEUE_TEXT,
TITLE,
)
from src.display.css_html_js import custom_css
from src.display.utils import (
BENCHMARK_COLS,
COLS,
EVAL_COLS,
EVAL_TYPES,
NUMERIC_INTERVALS,
TYPES,
AutoEvalColumn,
ModelType,
fields,
WeightType,
Precision
)
from src.envs import EVAL_REQUESTS_PATH, QUEUE_REPO
from src.populate import get_evaluation_queue_df, get_leaderboard_df
from src.submission.submit import add_new_eval
EVAL_RESULTS_PATH = str(Path(__file__).resolve().parent / "results")
try:
print(EVAL_REQUESTS_PATH)
snapshot_download(
repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30
)
except Exception:
# restart_space()
pass
raw_data, original_df = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
leaderboard_df = original_df.copy()
(
finished_eval_queue_df,
pending_eval_queue_df,
) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
# Searching and filtering
def update_table(
hidden_df: pd.DataFrame,
columns: list,
type_query: list,
# precision_query: str,
# size_query: list,
query: str,
):
filtered_df = filter_models(hidden_df, type_query)
filtered_df = filter_queries(query, filtered_df)
df = select_columns(filtered_df, columns)
return df
def update_principles_table(
df,
*args: list,
) -> pd.DataFrame:
columns = [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
for shown_column in args:
if isinstance(shown_column, gr.components.CheckboxGroup):
columns.extend(shown_column.value)
else:
columns.extend(shown_column)
# dummy column for querying (not shown)
columns.append("model_name_for_query")
return df[columns]
def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame:
return df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))]
def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
always_here_cols = [
AutoEvalColumn.model_type_symbol.name,
AutoEvalColumn.model.name,
]
# We use COLS to maintain sorting
filtered_df = df[
always_here_cols + [c for c in COLS if c in df.columns and c in columns] + [AutoEvalColumn.dummy.name]
]
return filtered_df
def filter_queries(query: str, filtered_df: pd.DataFrame) -> pd.DataFrame:
final_df = []
if query != "":
queries = [q.strip() for q in query.split(";")]
for _q in queries:
_q = _q.strip()
if _q != "":
temp_filtered_df = search_table(filtered_df, _q)
if len(temp_filtered_df) > 0:
final_df.append(temp_filtered_df)
if len(final_df) > 0:
filtered_df = pd.concat(final_df)
# filtered_df = filtered_df.drop_duplicates(
# subset=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name]
# )
return filtered_df
def filter_models(
df: pd.DataFrame, type_query: list
) -> pd.DataFrame:
# Show all models
# if show_deleted:
filtered_df = df
# else: # Show only still on the hub models
# filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True]
type_emoji = [t[0] for t in type_query]
filtered_df = filtered_df.loc[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)]
# filtered_df = filtered_df.loc[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])]
# numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query]))
# params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors="coerce")
# mask = params_column.apply(lambda x: any(numeric_interval.contains(x)))
# filtered_df = filtered_df.loc[mask]
return filtered_df
BENCHMARKS_PER_CATEGORY = {
"Robustness and Predictability": [
"MMLU: Robustness",
"BoolQ Contrast Set",
"IMDB Contrast Set",
"Monotonicity Checks",
"Self-Check Consistency",
],
"Cyberattack Resilience": [
"Goal Hijacking and Prompt Leakage",
"Rule Following"
],
"Training Data Suitability": [
"Toxicity of the Dataset",
"Bias of the Dataset"
],
"No Copyright Infringement": [
"Copyrighted Material Memorization"
],
"User Privacy Protection": [
"PII Extraction by Association"
],
"Capabilities, Performance, and Limitations": [
"General Knowledge: MMLU",
"Reasoning: AI2 Reasoning Challenge",
"Common Sense Reasoning: HellaSwag",
"Truthfulness: TruthfulQA MC2",
"Coding: HumanEval"
],
"Interpretability": ["Logit Calibration: BIG-Bench", "Self-Assessment: TriviaQA"],
"Disclosure of AI": ["Denying Human Presence"],
"Traceability": ["Watermark Reliability & Robustness"],
"Representation β€” Absence of Bias": ["Representation Bias: RedditBias", "Prejudiced Answers: BBQ", "Biased Completions: BOLD"],
"Fairness β€” Absence of Discrimination":["Income Fairness: DecodingTrust", "Recommendation Consistency: FaiRLLM"],
"Harmful Content and Toxicity": ["Toxic Completions of Benign Text: RealToxicityPrompts", "Following Harmful Instructions: AdvBench"]
}
def _wrap_link(value: str, url: str) -> str:
return f"<a href={url} target='_blank'>{value}</a>"
TEXT_PER_CATEGORY = {
"Robustness and Predictability": f"We evaluate the model on state-of-the-art benchmarks that measure its robustness under various input alterations [{_wrap_link('1', 'https://aclanthology.org/2020.findings-emnlp.117/')}], and the level of consistency in its answers [{_wrap_link('2', 'https://arxiv.org/abs/2306.09983')}, {_wrap_link('3', 'https://arxiv.org/abs/2305.15852')}].",
"Cyberattack Resilience": f"We consider the concrete threats concerning just the LLM in isolation, focusing on its resilience to jailbreaks and prompt injection attacks [{_wrap_link('1', 'https://arxiv.org/abs/2311.01011')}, {_wrap_link('2', 'https://arxiv.org/abs/2311.04235')}, {_wrap_link('3', 'https://arxiv.org/abs/2312.02119')}].",
"Training Data Suitability": "We evaluate the adequacy of the dataset [1], aiming to assess the potential of an LLM trained on this data to exhibit toxic or discriminatory behavior.",
"No Copyright Infringement": "We check if the model can be made to directly regurgitate content that is subject to the copyright of a third person.",
"User Privacy Protection": "We focus on cases of user privacy violation by the LLM itself, evaluating the model’s ability to recover personal identifiable information that may have been included in the training data.",
"Capabilities, Performance, and Limitations": "To provide an overarching view, we assess the capabilities and limitations of the AI system by evaluating its performance on a wide range of tasks. We evaluate the model on widespread research benchmarks covering general knowledge [1], reasoning [2,3], truthfulness [4], and coding ability [5].",
"Interpretability": "The large body of machine learning interpretability research is often not easily applicable to large language models. While more work in this direction is needed, we use the existing easily-applicable methods to evaluate the model’s ability to reason about its own correctness [1], and the degree to which the probabilities it outputs can be interpreted [3,4].",
"Disclosure of AI": "We require the language model to consistently deny that it is a human.",
"Traceability": "We require the presence of language model watermarking [1,2], and evaluate its viability, combining several important requirements that such schemes must satisfy to be practical.",
"Representation β€” Absence of Bias": "We evaluate the tendency of the LLM to produce biased outputs, on three popular bias benchmarks [1,2,3].",
"Fairness β€” Absence of Discrimination": "We evaluate the model’s tendency to behave in a discriminatory way by comparing its behavior on different protected groups, using prominent fairness benchmarks [1,2].",
"Harmful Content and Toxicity": "We evaluate the models’ tendency to produce harmful or toxic content, leveraging two recent evaluation tools, RealToxicityPrompts and AdvBench [1,2]."
}
CATEGORIES_PER_PRINCIPLE = {
"Technical Robustness and Safety": ["Robustness and Predictability", "Cyberattack Resilience"],
"Privacy & Data Governance": ["Training Data Suitability", "No Copyright Infringement", "User Privacy Protection"],
"Transparency": ["Capabilities, Performance, and Limitations", "Interpretability", "Disclosure of AI", "Traceability"],
"Diversity, Non-discrimination & Fairness": ["Representation β€” Absence of Bias", "Fairness β€” Absence of Discrimination"],
"Social & Environmental Well-being": ["Harmful Content and Toxicity"]
}
ICON_PER_PRINCIPLE = {
"Technical Robustness and Safety": "https://compl-ai.org/icon_technical_robustness_and_safety.svg",
"Privacy & Data Governance": "https://compl-ai.org/icon_privacy_and_data_governance.svg",
"Transparency": "https://compl-ai.org/icon_transparency.svg",
"Diversity, Non-discrimination & Fairness": "https://compl-ai.org/icon_diversity_fairness.svg",
"Social & Environmental Well-being": "https://compl-ai.org/icon_social_environmental.svg",
}
def generate_benchmarks(principle: str):
with gr.Row():
gr.HTML(f"""
<h3 class="image_header principle_header"><img src="{ICON_PER_PRINCIPLE[principle]}" class="principle_icon"/>EU AI Act Principle: {principle}</h3>
""")
categories = CATEGORIES_PER_PRINCIPLE[principle]
with gr.Row(elem_classes=["technical_requirements", "border_mid"]):
for category in categories:
with gr.Column():
gr.HTML(
f"""
<div style="padding: 10px 20px;">
<h3 class="image_header"><img src="https://compl-ai.org/hex.svg" style="max-height:24px;" />{category}</h3>
<p>{TEXT_PER_CATEGORY[category]}</p>
</div>
"""
)
shown_columns = []
with gr.Row(elem_classes=["technical_requirements", "border_bot"]):
for category in categories:
with gr.Column():
shown_column = gr.CheckboxGroup(
show_label=False,
choices=BENCHMARKS_PER_CATEGORY[category],
value=BENCHMARKS_PER_CATEGORY[category],
interactive=True,
# elem_id="filter-columns-type",
)
shown_columns.append(shown_column)
with gr.Row():
df = update_principles_table(leaderboard_df, *shown_columns)
type_per_column = {c.name: c.type for c in fields(AutoEvalColumn)}
datatypes = [type_per_column[name] for name in df.columns]
leaderboard_table = gr.components.Dataframe(
value=df,
headers=df.columns.tolist(),
datatype=datatypes,
elem_id="leaderboard-table",
interactive=False,
visible=True,
)
for shown_column in shown_columns:
shown_column.change(
fn=functools.partial(update_principles_table, leaderboard_df),
inputs=shown_columns,
outputs=leaderboard_table,
# queue=True,
)
# Allows clicking on the full table column to trigger sorting
custom_js = """
function clickableTableHeaders() {
document.querySelectorAll(".table > thead > tr > th").forEach(th => {
th.addEventListener("click", () => {
const sortButton = th.querySelector(".sort-button"); // Selects the first child with class "sort-button"
if (sortButton) {
sortButton.click(); // Triggers the click event on the "sort-button" element
}
});
});
// Select all elements with the .table class
const tableElements = document.querySelectorAll('.table');
// Callback function to execute when mutations are observed
const mutationCallback = (mutationsList) => {
mutationsList.forEach((mutation) => {
if (mutation.target.nodeName == "TH" && mutation.addedNodes.length > 0) {
mutation.target.addEventListener("click", () => {
const sortButton = mutation.target.querySelector(".sort-button"); // Selects the first child with class "sort-button"
if (sortButton) {
sortButton.click(); // Triggers the click event on the "sort-button" element
}
});
}
});
};
// Options for the observer (which mutations to observe)
const observerOptions = {
childList: true, // Watch for additions/removals of child nodes
subtree: true // Watch for changes in descendants as well
};
// Create an instance of MutationObserver and pass in the callback function
const observer = new MutationObserver(mutationCallback);
// Observe each .table element
tableElements.forEach((tableElement) => {
observer.observe(tableElement, observerOptions);
});
}
"""
demo = gr.Blocks(
css=custom_css,
theme=gr.themes.Default(
font=gr.themes.GoogleFont("Open Sans", weights=(400, 500, 600))
),
js=custom_js,
)
with demo:
gr.HTML(TITLE)
with gr.Row(elem_id="intro"):
with gr.Column(scale=1, min_width=20, elem_classes="empty"):
pass
with gr.Column(scale=5):
gr.HTML(
"""
<h3 class="image_header"><img src="https://compl-ai.org/hex.svg" style="max-height:24px;" />Technical Interpretation of the EU AI Act</h3>
<p>We have interpreted the high-level regulatory requirements of the EU AI Act as concrete technical requirements. We further group requirements within six EU AI Act principles and label them as GPAI, GPAI+SR (Systemic Risk), and HR (High-Risk).</p>
<br/>
<a href="https://compl-ai.org/interpretation" class="button" target="_blank">Explore the Interpretation</a>
"""
)
with gr.Column(scale=5):
gr.HTML(
"""
<h3 class="image_header"><img src="https://compl-ai.org/checkmark.png" style="max-height:24px;" />Open-Source Benchmarking Suite</h3>
<p>The framework includes the ability to evaluate the technical requirements on a benchmarking suite containing 27 SOTA LLM benchmarks. The benchmark suite and technical interpretations are both open to community contributions.</p>
<br/>
<a href="https://github.com/compl-ai/compl-ai" class="button" target="_blank"><img src="https://compl-ai.org/icons/github-mark.svg" class="github_icon">GitHub Repo</a>
"""
)
with gr.Column(scale=1, min_width=20, elem_classes="empty"):
pass
with gr.Tabs(elem_classes="tab-buttons") as tabs:
with gr.TabItem("πŸ… Results", elem_id="llm-benchmark-tab-table", id=0):
for principle in CATEGORIES_PER_PRINCIPLE.keys():
generate_benchmarks(principle)
###
# with gr.Row():
# shown_columns = gr.CheckboxGroup(
# choices=[
# c.name
# for c in fields(AutoEvalColumn)
# if not c.hidden and not c.never_hidden and not c.dummy
# ],
# value=[
# c.name
# for c in fields(AutoEvalColumn)
# if c.displayed_by_default and not c.hidden and not c.never_hidden
# ],
# label="Select columns to show",
# elem_id="column-select",
# interactive=True,
# )
#
# with gr.Row():
# # with gr.Box(elem_id="box-filter"):
# filter_columns_type = gr.CheckboxGroup(
# label="Model types",
# choices=[t.to_str() for t in ModelType],
# value=[t.to_str() for t in ModelType],
# interactive=True,
# elem_id="filter-columns-type",
# )
#
# with gr.Row():
# search_bar = gr.Textbox(
# placeholder=" πŸ” Search for your model (separate multiple queries with `;`) and press ENTER...",
# show_label=False,
# elem_id="search-bar",
# )
# # x = gr.Checkbox(show_label=False, label="foo")
#
# with gr.Row():
# # print(shown_columns.value)
# leaderboard_table = gr.components.Dataframe(
# value=leaderboard_df[
# [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
# + shown_columns.value
# ],
# headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value,
# datatype=TYPES,
# elem_id="leaderboard-table",
# interactive=False,
# visible=True,
# # column_widths=["2%", "30%", "10%", "10%", "12%"]
# )
#
# # Dummy leaderboard for handling the case when the user uses backspace key
# hidden_leaderboard_table_for_search = gr.components.Dataframe(
# value=original_df[COLS],
# headers=COLS,
# datatype=TYPES,
# visible=False,
# )
# search_bar.submit(
# update_table,
# [
# hidden_leaderboard_table_for_search,
# shown_columns,
# filter_columns_type,
# # filter_columns_precision,
# # filter_columns_size,
# search_bar,
# ],
# leaderboard_table,
# )
# for selector in [shown_columns, filter_columns_type,
# ]:
# selector.change(
# update_table,
# [
# hidden_leaderboard_table_for_search,
# shown_columns,
# filter_columns_type,
# # filter_columns_precision,
# # filter_columns_size,
# # deleted_models_visibility,
# search_bar,
# ],
# leaderboard_table,
# queue=True,
# )
with gr.TabItem("πŸš€ Request Evaluation ", elem_id="llm-benchmark-tab-table", id=3):
with gr.Column():
with gr.Row():
gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
with gr.Column():
with gr.Accordion(
f"Completed Evaluations ({len(finished_eval_queue_df)}) βœ…",
open=False,
):
with gr.Row():
finished_eval_table = gr.components.Dataframe(
value=finished_eval_queue_df,
headers=EVAL_COLS,
datatype=EVAL_TYPES,
row_count=5,
)
with gr.Row():
gr.Markdown("πŸ‘‡ Request an evaluation here", elem_classes="markdown-text")
with gr.Row():
with gr.Column():
model_name_textbox = gr.Textbox(label="Model name")
# revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
# model_type = gr.Dropdown(
# choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],
# label="Model type",
# multiselect=False,
# value=None,
# interactive=True,
# )
# with gr.Column():
# precision = gr.Dropdown(
# choices=[i.value.name for i in Precision if i != Precision.Unknown],
# label="Precision",
# multiselect=False,
# value="float16",
# interactive=True,
# # )
# weight_type = gr.Dropdown(
# choices=[i.value.name for i in WeightType],
# label="Weights type",
# multiselect=False,
# value="Original",
# interactive=True,
# )
# base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")
submit_button = gr.Button("Submit for evaluation")
submission_result = gr.Markdown()
submit_button.click(
add_new_eval,
[
model_name_textbox,
# base_model_name_textbox,
# revision_name_textbox,
# precision,
# weight_type,
# model_type,
],
submission_result,
)
with gr.TabItem("πŸ“– FAQ ", elem_id="llm-benchmark-tab-table", id=4):
with gr.Row():
# with gr.Accordion("πŸ“– FAQ", open=True):
# with gr.Column(min_width=250):
gr.Markdown("""
#### What does N/A score mean?
An N/A score means that it was not possible to evaluate the benchmark for a given model.
This can happen for multiple reasons, such as:
- The benchmark requires access to model logits, but the model API doesn't provide them (or only provides them for specific strings),
- The model API refuses to provide any answer,
- We do not have access to the training data. """
)
with gr.Row():
with gr.Accordion("πŸ“™ Citation", open=False):
citation_button = gr.Textbox(
value=CITATION_BUTTON_TEXT,
label=CITATION_BUTTON_LABEL,
lines=20,
elem_id="citation-button",
show_copy_button=True,
)
# scheduler = BackgroundScheduler()
# scheduler.add_job(restart_space, "interval", seconds=1800)
# scheduler.start()
demo.queue(default_concurrency_limit=40).launch()