import json from collections import defaultdict from dataclasses import dataclass, field, fields from functools import cached_property from pathlib import Path from typing import Literal import numpy as np import pandas as pd import gradio as gr from pandas import DataFrame from pandas.io.formats.style import Styler from content import * TASK_METRICS = { "arc": "acc_norm", "hellaswag": "acc_norm", "mmlu": "acc_norm", "truthfulqa": "mc2", } MODEL_TYPE_EMOJIS = { "pretrained": "🟢", "fine-tuned": "🔶", "instruction-tuned": "⭕", "RL-tuned": "🟦", } @dataclass class Result: model_name: str short_name: str model_type: Literal["pretrained", "fine-tuned", "instruction-tuned", "RL-tuned"] dutch_coverage: Literal["none", "pretrained", "fine-tuned"] num_parameters: int arc: float = field(default=0.0) average: float = field(default=0.0, init=False) hellaswag: float = field(default=0.0) mmlu: float = field(default=0.0) truthfulqa: float = field(default=0.0) num_parameters_kmb: str = field(init=False) def __post_init__(self): if self.model_type not in ["pretrained", "fine-tuned", "instruction-tuned", "RL-tuned"]: raise ValueError( f"Model type {self.model_type} must be one of 'pretrained', 'fine-tuned', 'instruction-tuned', 'RL-tuned'" ) if self.dutch_coverage not in ["none", "pretrained", "fine-tuned"]: raise ValueError(f"Dutch coverage {self.dutch_coverage} must be one of 'none', 'pretrained', 'fine-tuned'") field_names = {f.name for f in fields(self)} for task_name in TASK_METRICS: if task_name not in field_names: raise ValueError(f"Task name {task_name} not found in Result class fields so cannot create DataFrame") self.average = (self.arc + self.hellaswag + self.mmlu + self.truthfulqa) / 4 self.num_parameters_kmb = convert_number_to_kmb(self.num_parameters) @dataclass class ResultSet: results: list[Result] column_names: dict[str, str] = field(default_factory=dict) column_types: dict[str, str] = field(default_factory=dict) def __post_init__(self): if not self.column_names: # Order will be the order of the columns in the DataFrame self.column_names = { "short_name": "Model", "model_type": "T", "dutch_coverage": "🇳🇱", "num_parameters": "Size", "average": "Avg.", "arc": "ARC (25-shot)", "hellaswag": "HellaSwag (10-shot)️", "mmlu": "MMLU (5-shot)", "truthfulqa": "TruthfulQA (0-shot)", } self.column_types = { "Model": "markdown", "T": "str", "🇳🇱": "str", "Size": "str", "Avg.": "number", "ARC (25-shot)": "number", "HellaSwag (10-shot)️": "number", "MMLU (5-shot)": "number", "TruthfulQA (0-shot)": "number", } for column_type in self.column_types: if column_type not in set(self.column_names.values()): raise ValueError( f"Column names specified in column_types must be values in column_names." f" {column_type} not found." ) if "average" not in self.column_names: raise ValueError("Column names must contain 'average' column name") field_names = [f.name for f in fields(Result)] for column_name in self.column_names: if column_name not in field_names: raise ValueError(f"Column name {column_name} not found in Result class so cannot create DataFrame") @cached_property def df(self) -> DataFrame: data = [ { col_name: getattr(result, attr) for attr, col_name in self.column_names.items() } for result in self.results ] df = pd.DataFrame(data) df = df.sort_values(by=self.column_names["average"], ascending=False) return df @cached_property def styled_df(self) -> Styler: data = [ { col_name: (f"{result.short_name}") if attr == "short_name" else MODEL_TYPE_EMOJIS[result.model_type] if attr == "model_type" else getattr(result, attr) for attr, col_name in self.column_names.items() } for result in self.results ] df = pd.DataFrame(data) df = df.sort_values(by=self.column_names["average"], ascending=False) number_cols = [col for attr, col in self.column_names.items() if attr in TASK_METRICS or attr == "average"] styler = df.style.format("{:.2f}", subset=number_cols) def highlight_max(col): return np.where(col == np.nanmax(col.to_numpy()), "font-weight: bold;", None) styler = styler.apply(highlight_max, axis=0, subset=number_cols) num_params_col = self.column_names["num_parameters"] styler = styler.format(convert_number_to_kmb, subset=num_params_col) styler = styler.hide() return styler @cached_property def latex_df(self) -> Styler: number_cols = [col for attr, col in self.column_names.items() if attr in TASK_METRICS or attr == "average"] styler = self.df.style.format("{:.2f}", subset=number_cols) def highlight_max(col): return np.where(col == np.nanmax(col.to_numpy()), "font-weight: bold;", None) styler = styler.apply(highlight_max, axis=1, subset=number_cols) num_params_col = self.column_names["num_parameters"] styler = styler.format(convert_number_to_kmb, subset=num_params_col) styler = styler.hide() return styler def convert_number_to_kmb(number: int) -> str: """ Converts a number to a string with K, M or B suffix :param number: the number to convert :return: a string with the number and a suffix, e.g. "7B", rounded to one decimal """ if number >= 1_000_000_000: return f"{round(number / 1_000_000_000, 1)}B" elif number >= 1_000_000: return f"{round(number / 1_000_000, 1)}M" elif number >= 1_000: return f"{round(number / 1_000, 1)}K" else: return str(number) def collect_results() -> ResultSet: """ Collects results from the evals folder and returns a dictionary of results :return: a dictionary of results where the keys are typles of (model_name, language) and the values are dictionaries of the form {benchmark_name: performance_score} """ evals_dir = Path(__file__).parent.joinpath("evals") pf_overview = evals_dir.joinpath("models.json") if not pf_overview.exists(): raise ValueError( f"Overview file {pf_overview} not found. Make sure to generate it first with `generate_overview_json.py`." ) model_info = json.loads(pf_overview.read_text(encoding="utf-8")) model_results = {} for pfin in evals_dir.rglob("*.json"): data = json.loads(pfin.read_text(encoding="utf-8")) if "results" not in data: continue task_results = data["results"] short_name = pfin.stem.split("_", 2)[2].lower() if short_name not in model_results: model_results[short_name] = { "short_name": short_name, "model_name": model_info[short_name]["model_name"], "model_type": model_info[short_name]["model_type"], "dutch_coverage": model_info[short_name]["dutch_coverage"], "num_parameters": model_info[short_name]["num_parameters"], } for task_name, task_result in task_results.items(): task_name = task_name.rsplit("_", 1)[0] metric = TASK_METRICS[task_name] model_results[short_name][task_name] = task_result[metric] model_results = ResultSet([Result(**res) for short_name, res in model_results.items()]) return model_results with gr.Blocks() as demo: gr.HTML(TITLE) gr.Markdown(INTRO_TEXT) gr.Markdown(f"## Leaderboard\nOnly representative for the Dutch version (`*_nl`) of the benchmarks!") results = collect_results() gr.components.Dataframe( results.styled_df, headers=list(results.df.columns), datatype=[results.column_types[col] for col in results.df.columns], # To ensure same order as headers interactive=False, elem_id="leaderboard-table", ) with gr.Row(): with gr.Column(): modeltypes_str = "
".join([f"- {emoji}: {modeltype}" for modeltype, emoji in MODEL_TYPE_EMOJIS.items()]) gr.Markdown(f"Model types:
{modeltypes_str}") with gr.Column(): gr.Markdown( f"Language coverage ({results.column_names['dutch_coverage']}):" f"
- `none`: no explicit/deliberate Dutch coverage," f"
- `pretrained`: pretrained on Dutch data," f"
- `fine-tuned`: fine-tuned on Dutch data" ) with gr.Column(): metrics_str = "
".join([f"- {task}: `{metric}`" for task, metric in TASK_METRICS.items()]) gr.Markdown(f"Reported metrics:
{metrics_str}") gr.Markdown("## LaTeX") gr.Code(results.latex_df.to_latex(convert_css=True)) gr.Markdown(CREDIT, elem_classes="markdown-text") gr.Markdown(CITATION, elem_classes="markdown-text") if __name__ == "__main__": demo.launch()