app.py

import gradio as gr
import json
import os
import random
import datasets

from dipromats_evaluation_v2 import evaluate_results

# CONSTANTS

# Hugging Face datasets

DATASET_GOLD = "NLP-UNED/dipromats2024-t2_leaderboard-gold"
FILE_GOLD = 'gold_test.json'

DATASET_RESULTS = "NLP-UNED/dipromats2024-t2_leaderboard-results"
SPLIT_EN = 'results_en'
SPLIT_ES = 'results_es'
FEATURES_RESULTS = datasets.Features({
        "team_name": datasets.Value("string"),
        "run_id": datasets.Value("string"),
        "description": datasets.Value("string"),
        "lenient_f1": datasets.Value("float64"),
        "strict_f1": datasets.Value("float64"),
        "average_f1": datasets.Value("float64") })

EMPTY_RESULT={"team_name": [], "run_id": [], "description": [], "lenient_f1": [], "strict_f1": [], "average_f1": []}

# Before, you must create the Tokens in HF User Settings to give read and write access only to the datasets
try:
    from google.colab import userdata
    # Token must be copied and activated in Colab Secrets
    HF_TOKEN_GOLD = userdata.get('HF_DIPROMATS2024_T2_GOLD_TOKEN')
    HF_TOKEN_RESULTS = userdata.get('HF_DIPROMATS2024_T2_RESULTS_TOKEN')
except:
    # Assume running in HF Space
    # Tokens must be copied in Secrets under Space Settings
    HF_TOKEN_GOLD = os.getenv('HF_DIPROMATS2024_T2_GOLD_TOKEN')
    HF_TOKEN_RESULTS = os.getenv('HF_DIPROMATS2024_T2_RESULTS_TOKEN')

# LOAD DATASETS

# Load the Gold Standard data
# FILE_GOLD was uploaded directly through HF web, and the default split is train
dataset_gold = datasets.load_dataset(DATASET_GOLD, split='train', data_files=FILE_GOLD, token=HF_TOKEN_GOLD)
        
# Load the English dataset or create an empty one instead
try:
    dataset_en = datasets.load_dataset(DATASET_RESULTS, split=SPLIT_EN)
except Exception as e:
    print(f"Error loading English dataset: {e}. Creating it...")
    dataset_en = datasets.Dataset.from_dict(EMPTY_RESULT, features=FEATURES_RESULTS, split=SPLIT_EN)
    dataset_en.push_to_hub(DATASET_RESULTS, split=SPLIT_EN, token=HF_TOKEN_RESULTS)

# Load the Spanish dataset or create an empty one instead
try:
    dataset_es = datasets.load_dataset(DATASET_RESULTS, split=SPLIT_ES)
except Exception as e:
    print(f"Error loading Spanish dataset: {e}. Creating it...")
    dataset_es = datasets.Dataset.from_dict(EMPTY_RESULT, features=FEATURES_RESULTS, split=SPLIT_ES)
    dataset_es.push_to_hub(DATASET_RESULTS, split=SPLIT_ES, token=HF_TOKEN_RESULTS)

# AUX FUNCTIONS

# Función para convertir el dataset en tabla
def data_to_table(dataset):
    table_data = []
    for item in dataset:
        table_data.append([item.get("team_name", ""), item.get("run_id", ""),
                           item.get("lenient_f1", ""), item.get("strict_f1", ""), item.get("average_f1", "")])
    return table_data


# Función para subir los resultados al leaderboard
def update_leaderboard(lang, file_path, email, team_input, run_id, description, lenient_f1, strict_f1, average_f1):
    global dataset_en
    global dataset_es
    if lang == "en":
        dataset = dataset_en
    else:
        dataset = dataset_es

    warn = False
    if not email:
        gr.Warning("Email cannot be blank")
        warn=True
    if not team_input:
        gr.Warning("Team name cannot be blank")
        warn=True
    if not run_id:
        gr.Warning("Run ID cannot be blank")
        warn=True
    if not file_path:
        gr.Warning("File cannot be blank")
        warn=True
    if not description:
        gr.Warning("Description cannot be blank")
        warn=True

    if warn:
        return data_to_table(dataset_en), data_to_table(dataset_es), gr.Tabs(selected=1), gr.Button(visible=False), gr.Column(visible=True), team_input, run_id, description, email, file_path, lenient_f1, strict_f1, average_f1

    dataset = dataset.add_item({
        "team_name": team_input,
        "run_id": run_id,
        "description": description,
        "lenient_f1": lenient_f1,
        "strict_f1": strict_f1,
        "average_f1": average_f1
    })
    # Save change in database
    dataset.push_to_hub(DATASET_RESULTS, token=HF_TOKEN_RESULTS)

    # Update dataset in memory
    if lang == "en":
        dataset_en = dataset
    else:
        dataset_es = dataset

    #output: leaderboard_table, tabs, evaluate_button, submission_col, team_input, run_id, description_input, email_input, file_input, lenient_f1, strict_f1, average_f1
    return data_to_table(dataset_en), data_to_table(dataset_es), gr.Tabs(selected=0), gr.Button(visible=True), gr.Column(visible=False), "", "", "", "", None, None, None, None

# Función para procesar el archivo de resultados
def process_file(lang, file_path):
    global dataset_gold

    if not file_path:
        gr.Warning("File cannot be blank")
        return gr.Button(visible=True), gr.Row(visible=False), None, None, None

    with open(file_path, 'r') as f:
        test = json.load(f)

    try:
        results = evaluate_results(lang, dataset_gold, test)
        #print(results)
    except Exception as e:
        gr.Warning("Invalid JSON file or Incorrect Language")
        print(f"Error in function evaluate_results: {e}.")
        print(dataset_gold)
        return gr.Button(visible=True), gr.Row(visible=False), None, None, None

    lenient_f1 = results['lenient']['micro']['scores']['f1-score']
    strict_f1 = results['strict']['micro']['scores']['f1-score']
    average_f1 = (lenient_f1 + strict_f1) / 2

    return gr.Button(visible=False), gr.Row(visible=True), lenient_f1, strict_f1, average_f1


# Main

with gr.Blocks() as leaderboard:

    gr.Markdown(
        """
        # Dipromats 2024 Task 2 Leaderboard
        # Automatic Detection of Narratives from Diplomats of Major Powers
        These are the leaderboards for DIPROMATS 2024 Task 2 described in <a href=https://nlp.uned.es/dipromats2024>nlp.uned.es/dipromats2024</a>.
        The Gold Standard is not publicly available so LLMs cannot be contamined with them.
        However, you can submit your results here and get your system automatically evaluated.
        Then you will have the choice to submit your results to the leaderboard.
        """)
    with gr.Tabs() as tabs:

        # Tab English Leaderboard
        with gr.TabItem("English Leaderboard", id=0):
            gr.Markdown(
                """
                # English Leaderboard
                """)
            leaderboard_table_en = gr.Dataframe(headers=["Team", "Run ID", "Lenient F1", "Strict F1", "Average F1"],
                        value=data_to_table(dataset_en),
                        interactive=False)

        # Tab Spanish Leaderboard
        with gr.TabItem("Spanish Leaderboard", id=2):
            gr.Markdown(
                """
                # Spanish Leaderboard
                """)
            leaderboard_table_es = gr.Dataframe(headers=["Team", "Run ID", "Lenient F1", "Strict F1", "Average F1"],
                        value=data_to_table(dataset_es),
                        interactive=False)

        # Tab Evaluate
        with gr.TabItem("Evaluate your results", id=1):
            gr.Markdown(
                """
                # Upload your results and get evaluated
                Then you can decide to submit your results to the leaderboard or not.
                Make sure that you upload a file with the json format described in...
                """)
            with gr.Row():
                file_input = gr.File(label="Upload a JSON file", file_types=[".json"], type="filepath", file_count="single")
                with gr.Column():
                    lang = gr.Dropdown(label="Language", choices=["en", "es"], interactive=True)
                    evaluate_button = gr.Button("Evaluate")

            # System results table
            with gr.Row(visible=True):
                lenient_f1 = gr.Number(label="Lenient F1", interactive=False)
                strict_f1 = gr.Number(label="Strict F1", interactive=False)
                average_f1 = gr.Number(label="Average F1", interactive=False)

            # Submit to leaderboard
            with gr.Column(visible=False) as submission_col:
                with gr.Row():
                    with gr.Column():
                        with gr.Row():
                            team_input = gr.Textbox(label="Team Name")
                            run_id = gr.Textbox(label="Run ID")
                        email_input = gr.Textbox(label="Email (only for submission verification, it won't be shown)")
                    description_input = gr.Textbox(label="System description", lines=6)
                submit_button = gr.Button("Submit to leaderboard")

    evaluate_button.click(process_file,
                        inputs=[lang, file_input],
                        outputs=[evaluate_button, submission_col,lenient_f1, strict_f1, average_f1])

    submit_button.click(update_leaderboard,
                        inputs=[lang, file_input, email_input, team_input, run_id, description_input, lenient_f1, strict_f1, average_f1],
                        outputs=[leaderboard_table_en, leaderboard_table_es, tabs, evaluate_button, submission_col, team_input, run_id, description_input, email_input, file_input, lenient_f1, strict_f1, average_f1])

leaderboard.launch()