from time import sleep
import gradio as gr
import pandas as pd
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import snapshot_download
# from gradio_space_ci import enable_space_ci
from src.display.about import (
INTRODUCTION_TEXT,
LLM_BENCHMARKS_TEXT,
CITATION_BUTTON_LABEL,
CITATION_BUTTON_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,
Format
)
from src.envs import API, EVAL_RESULTS_PATH, RESULTS_REPO, REPO_ID, HF_TOKEN
from src.populate import get_leaderboard_df
# from src.tools.collections import update_collections
from src.tools.plots import (
create_metric_plot_obj,
create_plot_df,
create_scores_df,
)
# Start ephemeral Spaces on PRs (see config in README.md)
# enable_space_ci()
def restart_space():
API.restart_space(repo_id=REPO_ID, token=HF_TOKEN)
def init_space():
try:
print(EVAL_RESULTS_PATH)
snapshot_download(
repo_id=RESULTS_REPO,
local_dir=EVAL_RESULTS_PATH,
repo_type="dataset",
tqdm_class=None,
etag_timeout=30,
resume_download=True,
)
except Exception as e:
print(e)
sleep(180) # sleep 3 min
return init_space()
raw_data, original_df = get_leaderboard_df(
results_path=EVAL_RESULTS_PATH, cols=COLS, benchmark_cols=BENCHMARK_COLS
)
# update_collections(original_df.copy())
leaderboard_df = original_df.copy()
plot_df = create_plot_df(create_scores_df(raw_data))
return leaderboard_df, original_df, plot_df
leaderboard_df, original_df, plot_df = init_space()
# Searching and filtering
def update_table(
hidden_df: pd.DataFrame,
columns: list,
# type_query: list,
weight_precision_query: str,
activation_precision_query: str,
size_query: list,
hide_models: list,
format_query: list,
query: str,
):
filtered_df = filter_models(
df=hidden_df,
# type_query=type_query,
size_query=size_query,
weight_precision_query=weight_precision_query,
activation_precision_query=activation_precision_query,
hide_models=hide_models,
format_query=format_query,
)
filtered_df = filter_queries(query, filtered_df)
df = select_columns(filtered_df, columns)
return df
def load_query(request: gr.Request): # triggered only once at startup => read query parameter if it exists
query = request.query_params.get("query") or ""
return (
query,
query,
) # return one for the "search_bar", one for a hidden component that triggers a reload only if value has changed
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 = [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
dummy_col = [AutoEvalColumn.dummy.name]
# 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] + dummy_col]
return filtered_df
def filter_queries(query: str, filtered_df: pd.DataFrame):
"""Added by Abishek"""
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.weight_precision.name,
AutoEvalColumn.activation_precision.name,
AutoEvalColumn.revision.name,
]
)
return filtered_df
def filter_models(
df: pd.DataFrame,
# type_query: list,
size_query: list,
weight_precision_query: list,
activation_precision_query: list,
hide_models: list,
format_query: list,
) -> pd.DataFrame:
# Show all models
if "Private or deleted" in hide_models:
filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True]
else:
filtered_df = df
if "Contains a merge/moerge" in hide_models:
filtered_df = filtered_df[filtered_df[AutoEvalColumn.merged.name] == False]
if "MoE" in hide_models:
filtered_df = filtered_df[filtered_df[AutoEvalColumn.moe.name] == False]
if "Flagged" in hide_models:
filtered_df = filtered_df[filtered_df[AutoEvalColumn.flagged.name] == False]
# 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.weight_precision.name].isin(weight_precision_query + ["None"])]
filtered_df = filtered_df.loc[
df[AutoEvalColumn.activation_precision.name].isin(activation_precision_query + ["None"])
]
filtered_df = filtered_df.loc[df[AutoEvalColumn.format.name].isin(format_query)]
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
leaderboard_df = filter_models(
df=leaderboard_df,
# type_query=[t.to_str(" : ") for t in ModelType],
size_query=list(NUMERIC_INTERVALS.keys()),
weight_precision_query=[i.value.name for i in Precision],
activation_precision_query=[i.value.name for i in Precision],
hide_models=["Private or deleted", "Contains a merge/moerge", "Flagged"], # Deleted, merges, flagged, MoEs
format_query=[i.value.name for i in Format],
)
demo = gr.Blocks(css=custom_css)
with demo:
gr.HTML(TITLE)
gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
with gr.Tabs(elem_classes="tab-buttons") as tabs:
with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
with gr.Row():
with gr.Column():
with gr.Row():
search_bar = gr.Textbox(
placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...",
show_label=False,
elem_id="search-bar",
)
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():
hide_models = gr.CheckboxGroup(
label="Hide models",
choices=["Private or deleted", "Contains a merge/moerge", "Flagged"], #, "MoE"],
value=["Private or deleted", "Contains a merge/moerge", "Flagged"],
interactive=True,
)
with gr.Column(min_width=320):
# 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",
# )
filter_columns_weight_precision = gr.CheckboxGroup(
label="Weight Precision",
choices=[i.value.name for i in Precision],
value=[i.value.name for i in Precision],
interactive=True,
elem_id="filter-columns-weight-precision",
)
filter_columns_activation_precision = gr.CheckboxGroup(
label="Activation Precision",
choices=[i.value.name for i in Precision],
value=[i.value.name for i in Precision],
interactive=True,
elem_id="filter-columns-activation-precision",
)
filter_columns_size = gr.CheckboxGroup(
label="Model sizes (in billions of parameters)",
choices=list(NUMERIC_INTERVALS.keys()),
value=list(NUMERIC_INTERVALS.keys()),
interactive=True,
elem_id="filter-columns-size",
)
filter_format = gr.CheckboxGroup(
label="Format",
choices=[i.value.name for i in Format],
value=[i.value.name for i in Format],
interactive=True,
elem_id="filter-format",
)
leaderboard_table = gr.components.Dataframe(
value=leaderboard_df[
[c.name for c in fields(AutoEvalColumn) if c.never_hidden]
+ shown_columns.value
+ [AutoEvalColumn.dummy.name]
],
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%", "33%"]
)
# 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_weight_precision,
filter_columns_activation_precision,
filter_columns_size,
hide_models,
filter_format,
search_bar,
],
leaderboard_table,
)
# Define a hidden component that will trigger a reload only if a query parameter has been set
hidden_search_bar = gr.Textbox(value="", visible=False)
hidden_search_bar.change(
update_table,
[
hidden_leaderboard_table_for_search,
shown_columns,
# filter_columns_type,
filter_columns_weight_precision,
filter_columns_activation_precision,
filter_columns_size,
hide_models,
filter_format,
search_bar,
],
leaderboard_table,
)
# Check query parameter once at startup and update search bar + hidden component
demo.load(load_query, inputs=[], outputs=[search_bar, hidden_search_bar])
for selector in [
shown_columns,
# filter_columns_type,
filter_columns_weight_precision,
filter_columns_activation_precision,
filter_columns_size,
hide_models,
filter_format,
]:
selector.change(
update_table,
[
hidden_leaderboard_table_for_search,
shown_columns,
# filter_columns_type,
filter_columns_weight_precision,
filter_columns_activation_precision,
filter_columns_size,
hide_models,
filter_format,
search_bar,
],
leaderboard_table,
queue=True,
)
with gr.TabItem("📈 Metrics through time", elem_id="llm-benchmark-tab-table", id=4):
with gr.Row():
with gr.Column():
chart = create_metric_plot_obj(
plot_df,
[AutoEvalColumn.average.name],
title="Average of Top Scores and Human Baseline Over Time (from last update)",
)
gr.Plot(value=chart, min_width=500)
with gr.Column():
chart = create_metric_plot_obj(
plot_df,
BENCHMARK_COLS,
title="Top Scores and Human Baseline Over Time (from last update)",
)
gr.Plot(value=chart, min_width=500)
with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
with gr.Row():
with gr.Accordion("📙 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) # restarted every 3h
scheduler.start()
demo.queue(default_concurrency_limit=40).launch()