app.py

import os
import json
import glob
from collections import defaultdict
import pandas as pd
import gradio as gr
from content import *
from css import *
import glob

AFRIMMLU_DIRECT = "afrimmlu_direct"
AFRIMMLU_TRANSLATE = "afrimmlu_translate"
AFRIXNLI_DIRECT = "afrixnli_direct"
AFRIXNLI_TRANSLATE = "afrixnli_translate"


BENCHMARKS = [AFRIMMLU_DIRECT, AFRIMMLU_TRANSLATE, AFRIXNLI_DIRECT, AFRIXNLI_TRANSLATE]

METRICS = ["acc", "acc_stderr", "f1"]

LANGS = ['amh', 'eng', 'ewe', 'fra', 'hau', 'ibo', 'kin', 'lin', 'lug', 'orm', 'sna', 'sot', 'swa', 'twi', 'wol', 'xho', 'yor', 'zul']

LANG_NAME = {
    'amh': 'Amharic',
    'eng': 'English',
    'ewe': 'Ewe',
    'fra': 'French',
    'hau': 'Hausa',
    'ibo': 'Igbo',
    'kin': 'Kinyarwanda',
    'lin': 'Lingala',
    'lug': 'Luganda',
    'orm': 'Oromo',
    'sna': 'Shona',
    'sot': 'Sotho',
    'swa': 'Swahili',
    'twi': 'Twi',
    'wol': 'Wolof',
    'xho': 'Xhosa',
    'yor': 'Yoruba',
    'zul': 'Zulu'
}


def collect_results():
    performance_dict = defaultdict(dict)
    pretrained_models = set()
    for file in glob.glob('evals/*/*.json'):
        with open(file, 'r') as f:
            data = json.load(f)
        if 'results' not in data:
            continue
        if 'config' not in data:
            continue
        results = data['results']
        config = data['config']
        if 'model_args' not in config:
            continue

        model_args = config['model_args'].split(',')
        pretrained = [x for x in model_args if x.startswith('pretrained=')]
        if len(pretrained) != 1:
            continue
        pretrained = pretrained[0].split('=')[1]
        pretrained = pretrained.split('/')[-1]
        pretrained_models.add(pretrained)

        for lang_task, perfs in results.items():
            lang_task = lang_task.split('_')
            lang = lang_task[-1]
            task = '_'.join(lang_task[:-1])

            assert task in BENCHMARKS

            if lang and task:
                metric = METRICS[BENCHMARKS.index(task)-1]
                p = round(perfs[metric] * 100, 1)
                performance_dict[(pretrained, lang)][task] = p
    return performance_dict, pretrained_models


def get_leaderboard_df(performance_dict, pretrained_models):

    df = list()
    for (pretrained, lang), perfs in performance_dict.items():
        lang_name = LANG_NAME[lang]
        afrimmlu_direct_perf = perfs.get(AFRIMMLU_DIRECT, 0.0)
        afrimmlu_translate_perf = perfs.get(AFRIMMLU_TRANSLATE, 0.0)
        afrixnli_direct_perf = perfs.get(AFRIXNLI_DIRECT, 0.0)
        afrixnli_translate_perf = perfs.get(AFRIXNLI_TRANSLATE, 0.0)

        # if afrimmlu_direct_perf * afrimmlu_translate_perf * afrixnli_direct_perf * afrixnli_translate_perf == 0:
        #     continue

        average_divide = [1 if afrimmlu_direct_perf else 0, 1 if afrimmlu_translate_perf else 0, 1 if afrixnli_direct_perf else 0, 1 if afrixnli_translate_perf else 0]

        avg = round((afrimmlu_direct_perf + afrimmlu_translate_perf + afrixnli_direct_perf + afrixnli_translate_perf) / sum(average_divide), 1)
        notes = ' '.join([pretrained, lang_name])
        row = [pretrained, lang_name, lang, avg, afrimmlu_direct_perf, afrimmlu_translate_perf, afrixnli_direct_perf, afrixnli_translate_perf, notes]
        df.append(row)

    df = pd.DataFrame.from_records(df, columns=COLS)
    df = df.sort_values(by=[LANG_COL, AVERAGE_COL], ascending=False)
    df = df[COLS]

    return df


def search_table(df, query):
    filtered_df = df[df[NOTES_COL].str.contains(query, case=False)]
    return filtered_df



MODEL_COL = "Model"
LANG_COL = "Language"
CODE_COL = "Code"
AVERAGE_COL = "Average"
AFRIMMLU_DIRECT_COL = "AfriMMLU Direct (0-Shot)"
AFRIMMLU_TRANSLATE_COL = "AfriMMLU Translate (0-Shot)"
AFRIXNLI_DIRECT_COL = "AfriXNLI Direct (0-Shot)"
AFRIXNLI_TRANSLATE_COL = "AfriXNLI Translate (0-Shot)"
NOTES_COL = "Notes"  # For search only

COLS = [MODEL_COL, LANG_COL, CODE_COL, AVERAGE_COL, AFRIMMLU_DIRECT_COL, AFRIMMLU_TRANSLATE_COL, AFRIXNLI_DIRECT_COL, AFRIXNLI_TRANSLATE_COL, NOTES_COL]
TYPES = ["str", "str", "str", "number", "number", "number", "number", "number", "str"]

args = collect_results()
original_df = get_leaderboard_df(*args)

demo = gr.Blocks(css=CUSTOM_CSS)
with demo:
    gr.HTML(TITLE)
    gr.Markdown(INTRO_TEXT, elem_classes="markdown-text")
    gr.Markdown(HOW_TO, elem_classes="markdown-text")

    with gr.Group():
        search_bar = gr.Textbox(
            placeholder="Search models and languages...", show_label=False, elem_id="search-bar"
        )

        leaderboard_table = gr.components.Dataframe(
            value=original_df,
            headers=COLS,
            datatype=TYPES,
            # max_rows=5,
            elem_id="leaderboard-table",
        )

        # # Dummy leaderboard for handling the case when the user uses backspace key
        hidden_leaderboard_table_for_search = gr.components.Dataframe(
            value=original_df, 
            headers=COLS, 
            datatype=TYPES, 
            # max_rows=5, 
            visible=False
        )

        search_bar.change(
            search_table,
            [hidden_leaderboard_table_for_search, search_bar],
            leaderboard_table,
        )

    gr.Markdown(CREDIT, elem_classes="markdown-text")
    gr.Markdown(CITATION, elem_classes="markdown-text")

demo.launch()