server.py

import copy
import glob
import json
import os
# Necessary for `requests`. Without set correct path or empty string it fails during process HTTPS connection with this: [Errno 101] Network is unreachable
if os.path.exists("/etc/ssl/certs/ca-certificates.crt"):
    os.environ["CURL_CA_BUNDLE"] = "/etc/ssl/certs/ca-certificates.crt"
    os.environ["REQUESTS_CA_BUNDLE"] = "/etc/ssl/certs/ca-certificates.crt"
else:
    os.environ["CURL_CA_BUNDLE"] = ""
    os.environ["REQUESTS_CA_BUNDLE"] = ""
print(f"{os.environ.get('CURL_CA_BUNDLE') = }")
print(f"{os.environ.get('REQUESTS_CA_BUNDLE') = }")

import hashlib
import time
import requests
from collections import namedtuple
from xml.sax.saxutils import escape as xmlEscape, quoteattr as xmlQuoteAttr
from threading import Lock

import gradio as gr
import pandas as pd
from huggingface_hub import HfApi, snapshot_download

from compare_significance import SUPPORTED_METRICS

VISIBLE_METRICS = SUPPORTED_METRICS + ["macro_f1"]

api = HfApi()

HF_RESULTS_DATASET = os.environ.get("HF_RESULTS_DATASET")  # <HF_RESULTS_DATASET> ::= <owner> "/" <dataset name>; e.g. CZLC/LLM_benchmark_data
HF_TOKEN = os.environ.get("HF_TOKEN")

TASKS_METADATA_PATH = "./tasks_metadata.json"

MARKDOWN_SPECIAL_CHARACTERS = {
    "#": "&#35;",  # for usage in xml.sax.saxutils.escape as entities must be first
    "\\": "&#92;",
    "`": "&#96;",
    "*": "&#42;",
    "_": "&#95;",
    "{": "&#123;",
    "}": "&#125;",
    "[": "&#91;",
    "]": "&#93;",
    "(": "&#40;",
    ")": "&#41;",
    "+": "&#43;",
    "-": "&#45;",
    ".": "&#46;",
    "!": "&#33;",
    "=": "&#61;",
    "|": "&#124;"
}

def xmlAndMarkdownEscape(text):
    return xmlEscape(text, MARKDOWN_SPECIAL_CHARACTERS)

class CheckSignificanceError(Exception):
    pass

def check_significance_send_task(model_a_path, model_b_path):
    url = 'https://czechllm.fit.vutbr.cz/benczechmark-leaderboard/compare_significance/'

    # prepare and send request
    with (
        open(model_a_path, 'rb') as model_a_fp,
        open(model_b_path, 'rb') as model_b_fp,
    ):
        files = {
            'model_a': model_a_fp,
            'model_b': model_b_fp,
        }
        response = requests.post(url, files=files, timeout=60 * 5)

    # check response
    if response.status_code == 202:
        result_url = response.url
        #task_id = response.json()['task_id']
    elif response.status_code == 429:
        raise CheckSignificanceError('Server is too busy. Please try again later.')
    else:
        raise CheckSignificanceError(f'Failed to submit task. Status code: {response.status_code}')

    return result_url

def check_significance_wait_for_result(result_url):
    while True:
        response = requests.get(result_url, timeout=60 * 5)
        if response.status_code == 200:
            result = response.json()
            break
        elif response.status_code == 202:
            time.sleep(5)
        else:
            raise CheckSignificanceError(f'Failed to get result. Status code: {response.status_code}')

    if result["state"] == "COMPLETED":
        return result['result']
    else:
        raise CheckSignificanceError(result['result']['error'])

def check_significance(model_a_path, model_b_path):
    result_url = check_significance_send_task(model_a_path, model_b_path)
    result = check_significance_wait_for_result(result_url)
    return result

pre_submit_lock = Lock()

class _ReadLock:
    def __init__(self, lock):
       self._lock = lock
       self.reading = 0
    
    def __enter__(self):
        with self._lock:
            self.reading += 1
    
    def __exit__(self, exc_type, exc_value, traceback):
        with self._lock:
            self.reading -= 1

class ReadWriteLock:
    """
    Zámek, který ověří, že nikdo nečte když se zapisuje a že zapisuje pouze jeden
    """
    
    def __init__(self):
       self._lock = Lock()
       self.ro = _ReadLock(self._lock)
       self.rw = self
    
    def __enter__(self):
        self._lock.acquire()
        while True:
            reading = self.ro.reading
            if reading > 0:
                self._lock.release()
                time.sleep(1)
                self._lock.acquire()
            elif reading < 0:
                self._lock.release()
                raise RuntimeError()
            else:
                return
    
    def __exit__(self, exc_type, exc_value, traceback):
        self._lock.release()

class LeaderboardServer:
    def __init__(self):
        self.SERVER_ADDRESS = HF_RESULTS_DATASET
        self.REPO_TYPE = "dataset"
        self.TASKS_METADATA = json.load(open(TASKS_METADATA_PATH))
        self.TASKS_CATEGORIES = {self.TASKS_METADATA[task]["category"] for task in self.TASKS_METADATA}
        self.TASKS_CATEGORY_OVERALL = "Overall"
        self.CATEGORY_TO_TASK_ABBREVIATION_TO_DETAILS = self._prepare_category_to_task_abbr_to_details()
        self.MAX_LENGTH_OF_MODEL_TITLE = 28
        self.DIR_DATAFRAMES_CSV = "./dataframes_csv"
        
        self.var_lock = ReadWriteLock()
        self.submission_ids = set()
        self.submission_id_to_file = {}  # Map submission ids to file paths
        self.submission_id_to_model_title = {}
        self.submission_id_to_data = {}  # Only data (results and metadata) using by leaderboard
        self.tournament_results = None
        self.leaderboard_dataframes = {}  # For each category
        self.tournament_dataframes = {}  # For each submission_id and category
        
        self.leaderboard_dataframes_csv = {}  # For each category
        self.tournament_dataframes_csv = {}  # For each submission_id and category
        
        self.results_dataset_local_snapshot_lock = ReadWriteLock()
        self.results_dataset_local_snapshot = None
        
        self.pre_submit_lock = pre_submit_lock
        self.pre_submit = None
        
        self.update_leaderboard()
        self.results_dataset_integrity_check()  # Check integrity of the results dataset after (re)start Hugging Face Space

    def update_leaderboard(self):
        with self.results_dataset_local_snapshot_lock.rw:
            self.results_dataset_local_snapshot = snapshot_download(
                self.SERVER_ADDRESS,
                repo_type=self.REPO_TYPE,
                token=HF_TOKEN,
                local_dir="./",
            )
        
        self.fetch_existing_models()
        
        tournament_results = self.load_tournament_results()
        with self.var_lock.rw:
            self.tournament_results = tournament_results
        
        categories = [self.TASKS_CATEGORY_OVERALL] + sorted(self.TASKS_CATEGORIES)
        
        leaderboard_dataframes = {
            category: self._get_leaderboard(category=category)
            for category in categories
        }
        
        with self.var_lock.ro:
            submission_ids = self.submission_ids
        
        tournament_dataframes = {
            submission_id: {
                category: self._get_model_tournament_table(submission_id, category)
                for category in categories
            }
            for submission_id in submission_ids
        }
        
        with self.var_lock.rw:
            self.leaderboard_dataframes = leaderboard_dataframes
            self.tournament_dataframes = tournament_dataframes
        
        leaderboard_dataframes_csv = {
            category: self._dataframe_to_csv(self._get_leaderboard(category=category, to_csv=True), f"Leaderboard - {category}.csv")
            for category in categories
        }
        
        with self.var_lock.ro:
            tournament_dataframes_csv = {
                submission_id: {
                    category: self._dataframe_to_csv(
                        self._get_model_tournament_table(submission_id, category, to_csv=True),
                        f"Tournament table - {self.submission_id_to_data[submission_id]['metadata']['model_name'][:self.MAX_LENGTH_OF_MODEL_TITLE].replace('/', '_')} - {category}.csv",
                    )
                    for category in sorted(self.TASKS_CATEGORIES)
                }
                for submission_id in submission_ids
            }
        
        with self.var_lock.rw:
            self.leaderboard_dataframes_csv = leaderboard_dataframes_csv
            self.tournament_dataframes_csv = tournament_dataframes_csv

    def load_tournament_results(self):
        with self.results_dataset_local_snapshot_lock.ro:
            metadata_rank_paths = os.path.join(self.results_dataset_local_snapshot, "tournament.json")
            if not os.path.exists(metadata_rank_paths):
                return {}
            with open(metadata_rank_paths) as ranks_file:
                results = json.load(ranks_file)
            return results

    def _prepare_category_to_task_abbr_to_details(self):
        tasks_per_category = {}
        for task in self.TASKS_METADATA:
            task_category = self.TASKS_METADATA[task]["category"]
            tasks_per_category.setdefault(task_category, list()).append(task)
        
        category2abbreviation2name = {self.TASKS_CATEGORY_OVERALL: {}}
        for category, tasks in tasks_per_category.items():
            abbreviation2name = {
                self.TASKS_METADATA[t]["abbreviation"]: (
                    self.TASKS_METADATA[t]["abbreviation"],
                    self.TASKS_METADATA[t]["name"],
                    self.TASKS_METADATA[t]["source_url"],
                )
                for t in tasks
            }
            sorted_abbreviation2name = dict.fromkeys(sorted(abbreviation2name.keys()))
            sorted_abbreviation2name.update(abbreviation2name)
            category2abbreviation2name[category] = sorted_abbreviation2name
            category2abbreviation2name[self.TASKS_CATEGORY_OVERALL].update(sorted_abbreviation2name)
        
        abbreviation2name = category2abbreviation2name[self.TASKS_CATEGORY_OVERALL]
        sorted_abbreviation2name = dict.fromkeys(sorted(abbreviation2name.keys()))
        sorted_abbreviation2name.update(abbreviation2name)
        category2abbreviation2name[self.TASKS_CATEGORY_OVERALL] = sorted_abbreviation2name
        
        return category2abbreviation2name

    def fetch_existing_models(self):
        # Models data
        submission_ids = set()
        submission_id_to_file = {}
        submission_id_to_model_title = {}
        submission_id_to_data = {}
        
        with self.results_dataset_local_snapshot_lock.ro:
            for submission_file in glob.glob(os.path.join(self.results_dataset_local_snapshot, "data") + "/*.json"):
                data = json.load(open(submission_file))
                metadata = data.get('metadata')
                if metadata is None:
                    continue
                submission_id = metadata["submission_id"]
                
                submission_ids.add(submission_id)
                submission_id_to_file[submission_id] = submission_file
                submission_id_to_model_title[submission_id] = metadata["team_name"] + "/" + metadata["model_name"]
                submission_id_to_data[submission_id] = {"results": data["results"], "metadata": metadata}
        
        with self.var_lock.rw:
            self.submission_ids = submission_ids
            self.submission_id_to_file = submission_id_to_file
            self.submission_id_to_model_title = submission_id_to_model_title
            self.submission_id_to_data = submission_id_to_data

    def results_dataset_integrity_check(self):
        """
        Zkontroluje, že:
        - všechny modely byly v duelu se všemi
        -- pokud ne, znemožní potvrzení nových submitů a udělá zbývající zápasy
        -- kontroluje soubory v adresáři "/data" a soubor "tournament.json"
        - v souboru "tournament.json" není `submission_id`, které by nemělo soubor v adresáři "/data"
        """
        
        while True:
            with self.pre_submit_lock:
                if self.pre_submit == None:
                    gr.Info('Checking integrity...', duration=15)
                    self.update_leaderboard()
                    
                    with self.var_lock.ro:
                        # Is every `submission_id` in results known?
                        if self.tournament_results.keys() - self.submission_ids != set():
                            pass
                        # Was every `submission_id` in some match?
                        elif self.submission_ids - self.tournament_results.keys() != set():
                            pass
                        # Are all competitors known?
                        elif any(
                            self.tournament_results[submission_id].keys() - self.submission_ids != set()
                            for submission_id in self.submission_ids
                        ):
                            pass
                        # Has had every `submission_id` match with all competitors?
                        elif any(
                            self.submission_ids - self.tournament_results[submission_id].keys() != set()
                            for submission_id in self.submission_ids
                        ):
                            pass
                        else:
                            break

                    gr.Info('Running tournament...', duration=15)
                    with self.var_lock.rw:
                        self.tournament_results = {}
                        submission_ids_backup = self.submission_ids
                        self.submission_ids = set()
                    for submission_id in submission_ids_backup:
                        with self.var_lock.ro:
                            file = self.submission_id_to_file[submission_id]
                            tournament_results = self.start_tournament(submission_id, file)
                        with self.var_lock.rw:
                            self.tournament_results = tournament_results
                            self.submission_ids.add(submission_id)
                    
                    gr.Info('Uploading tournament results...', duration=5)
                    if self.tournament_results:
                        self._upload_tournament_results(self.tournament_results)
                    
                    break
            gr.Info("Waiting in queue...", duration=5)
            time.sleep(10)
        
        gr.Info('Integrity of the results dataset is checked', duration=5)

    @staticmethod
    def _model_tournament_table_highlight_true_and_false(x):
        df_css = x.copy()
        for c in df_css:
          for i in range(len(df_css.index)):
            if x[c].iloc[i] == True or ">true<" in str(x[c].iloc[i]).lower():
              df_css[c].iloc[i] = 'background-color: rgba(0, 255, 0, 0.1);'
            elif x[c].iloc[i] == False or ">false<" in str(x[c].iloc[i]).lower():
              df_css[c].iloc[i] = 'background-color: rgba(255, 0, 0, 0.1);'
            else:
              df_css[c].iloc[i] = ''
        return df_css 

    def get_model_tournament_table_csv(self, submission_id, category):
        with self.var_lock.ro:
            return self.tournament_dataframes_csv[submission_id][category]

    def get_model_tournament_table(self, submission_id, category):
        with self.var_lock.ro:
            return copy.copy(self.tournament_dataframes[submission_id][category])

    def _get_model_tournament_table(self, submission_id, category, to_csv=False):
        if category == self.TASKS_CATEGORY_OVERALL:
            return None
        
        model_tournament_table = []
        
        with self.var_lock.ro:
            for competitor_id in self.tournament_results[submission_id].keys() - {submission_id}: # without self
                data = self.submission_id_to_data[competitor_id]
                
                match_results = {}
                for task in self.tournament_results[submission_id][competitor_id]:
                    task_category = self.TASKS_METADATA[task]["category"]
                    if task_category == category:
                        if to_csv:
                            match_results[task] = self.tournament_results[submission_id][competitor_id][task]["significant"]
                        else:
                            match_task_result_details = dict.fromkeys(["significant", "p_value"])  # order has impact to sorting DataFrame
                            match_task_result_details.update(copy.deepcopy(self.tournament_results[submission_id][competitor_id][task]))
                            match_task_result_details["significant"] = str(match_task_result_details["significant"]).lower()  # originaly bool
                            match_task_result_significant = match_task_result_details["significant"]
                            match_task_result_details = "\n".join(f"{k}: {v}" for k, v in match_task_result_details.items())
                            match_results[task] = f'<abbr title={xmlQuoteAttr(match_task_result_details)}>{match_task_result_significant}</abbr>'
                
                model_link = data["metadata"]["link_to_model"]
                model_title = data["metadata"]["team_name"] + "/" + data["metadata"]["model_name"]
                if to_csv:
                    match_results["model"] = model_title
                    match_results["link_to_model"] = model_link
                else:
                    model_title_abbr_team_name = self.abbreviate(data["metadata"]["team_name"], self.MAX_LENGTH_OF_MODEL_TITLE)
                    model_title_abbr_model_name = self.abbreviate(data["metadata"]["model_name"], self.MAX_LENGTH_OF_MODEL_TITLE)
                    model_title_abbr_html = f'<div style="font-size: 10px;">{xmlAndMarkdownEscape(model_title_abbr_team_name)}</div>{xmlAndMarkdownEscape(model_title_abbr_model_name)}'
                    match_results["model"] = f'<a href={xmlQuoteAttr(model_link)} title={xmlQuoteAttr(model_title)}>{model_title_abbr_html}</a>'
                
                model_tournament_table.append(match_results)
            
            dataframe = pd.DataFrame.from_records(model_tournament_table)
            
            extra_attributes_map_word_to_header = {
                "model": "Competitor",
                "link_to_model": "Link to model",
            }
            first_attributes = [
                "model",
                "link_to_model",
            ]
            df_order = [
                key
                for key in dict.fromkeys(
                    first_attributes
                    + sorted(
                        list(self.TASKS_METADATA.keys())
                        + list(dataframe.columns)
                    )
                ).keys()
                if key in dataframe.columns
            ]
            dataframe = dataframe[df_order]
            attributes_map_word_to_header = {key: value["abbreviation"] for key, value in self.TASKS_METADATA.items()}
            attributes_map_word_to_header.update(extra_attributes_map_word_to_header)
            dataframe = dataframe.rename(
                columns=attributes_map_word_to_header
            )
            if not to_csv:
                dataframe = dataframe.style.apply(self._model_tournament_table_highlight_true_and_false, axis=None)
            return dataframe

    def _dataframe_to_csv(self, dataframe, filename):
        try:
            if not os.path.isdir(self.DIR_DATAFRAMES_CSV):
                os.mkdir(self.DIR_DATAFRAMES_CSV)
        except FileExistsError:
            pass
        
        filepath = os.path.join(self.DIR_DATAFRAMES_CSV, filename)
        dataframe.to_csv(filepath, index=False)
        return filepath

    def get_leaderboard_csv(self, category=None):
        category = category if category else self.TASKS_CATEGORY_OVERALL
        with self.var_lock.ro:
            return self.leaderboard_dataframes_csv[category]

    def get_leaderboard(self, pre_submit=None, category=None):
        if pre_submit == None:
            category = category if category else self.TASKS_CATEGORY_OVERALL
            with self.var_lock.ro:
                return copy.copy(self.leaderboard_dataframes[category])
        else:
            return self._get_leaderboard(pre_submit=pre_submit, category=category)

    def _get_leaderboard(self, pre_submit=None, category=None, to_csv=False):
        with self.var_lock.ro:
            tournament_results = pre_submit.tournament_results if pre_submit else self.tournament_results
            category = category if category else self.TASKS_CATEGORY_OVERALL

            if len(tournament_results) == 0:
                return pd.DataFrame(columns=['No submissions yet'])
            else:
                processed_results = []
                for submission_id in tournament_results.keys():
                    if submission_id not in self.submission_id_to_data:
                        if pre_submit and submission_id == pre_submit.submission_id:
                            data = json.load(open(pre_submit.file))
                        else:
                            raise gr.Error(f"Internal error: Submission [{submission_id}] not found")
                    else:
                        data = self.submission_id_to_data[submission_id]
                    
                    if submission_id != data["metadata"]["submission_id"]:
                        raise gr.Error(f"Proper submission [{submission_id}] not found")

                    local_results = {}
                    win_score = {}
                    visible_metrics_map_word_to_header = {}
                    for task in self.TASKS_METADATA.keys():
                        
                        task_category = self.TASKS_METADATA[task]["category"]
                        if category not in (self.TASKS_CATEGORY_OVERALL, task_category):
                            continue
                        else:
                            # tournament_results
                            num_of_competitors = 0
                            num_of_wins = 0
                            for competitor_id in tournament_results[submission_id].keys() - {submission_id}: # without self
                                num_of_competitors += 1
                                if tournament_results[submission_id][competitor_id][task]["significant"]:
                                    num_of_wins += 1
                            task_score = num_of_wins / num_of_competitors * 100 if num_of_competitors > 0 else 100
                            win_score.setdefault(task_category, []).append(task_score)
                            
                            if category == task_category:
                                local_results[task] = task_score
                                for metric in VISIBLE_METRICS:
                                    visible_metrics_map_word_to_header[task + "_" + metric] = self.TASKS_METADATA[task]["abbreviation"] + " " + metric
                                    metric_value = data['results'][task].get(metric)
                                    if metric_value is not None:
                                        local_results[task + "_" + metric] = metric_value if metric == "word_perplexity" else metric_value * 100
                                        break  # Only the first metric of every task
                    
                    
                    for c in win_score:
                        win_score[c] = sum(win_score[c]) / len(win_score[c])
                    
                    if category == self.TASKS_CATEGORY_OVERALL:
                        for c in win_score:
                            local_results[c] = win_score[c]
                        local_results["average_score"] = sum(win_score.values()) / len(win_score)
                    else:
                        local_results["average_score"] = win_score[category]
                    
                    model_link = data["metadata"]["link_to_model"]
                    model_title = data["metadata"]["team_name"] + "/" + data["metadata"]["model_name"]
                    if to_csv:
                        local_results["model"] = model_title
                        local_results["link_to_model"] = model_link
                    else:
                        model_title_abbr_team_name = self.abbreviate(data["metadata"]["team_name"], self.MAX_LENGTH_OF_MODEL_TITLE)
                        model_title_abbr_model_name = self.abbreviate(data["metadata"]["model_name"], self.MAX_LENGTH_OF_MODEL_TITLE)
                        model_title_abbr_html = f'<div style="font-size: 10px;">{xmlAndMarkdownEscape(model_title_abbr_team_name)}</div>{xmlAndMarkdownEscape(model_title_abbr_model_name)}'
                        local_results["model"] = f'<a href={xmlQuoteAttr(model_link)} title={xmlQuoteAttr(model_title)}>{model_title_abbr_html}</a>'
                    
                    release = data["metadata"].get("submission_timestamp")
                    release = time.strftime("%Y-%m-%d", time.gmtime(release)) if release else "N/A"
                    local_results["release"] = release
                    local_results["model_type"] = data["metadata"]["model_type"]
                    local_results["parameters"] = data["metadata"]["parameters"]
                    
                    if pre_submit and submission_id == pre_submit.submission_id:
                        processed_results.insert(0, local_results)
                    else:
                        processed_results.append(local_results)
                dataframe = pd.DataFrame.from_records(processed_results)
                
                extra_attributes_map_word_to_header = {
                    "model": "Model",
                    "release": "Release",
                    "average_score": "Average ⬆️",
                    "team_name": "Team name",
                    "model_name": "Model name",
                    "model_type": "Type",
                    "parameters": "# θ (B)",
                    "input_length": "Input length (# tokens)",
                    "precision": "Precision",
                    "description": "Description",
                    "link_to_model": "Link to model",
                }
                first_attributes = [
                    "model",
                    "link_to_model",
                    "release",
                    "model_type",
                    "parameters",
                    "average_score",
                ]
                df_order = [
                    key
                    for key in dict.fromkeys(
                        first_attributes
                        + sorted(
                            list(self.TASKS_METADATA.keys())
                            + list(dataframe.columns)
                        )
                    ).keys()
                    if key in dataframe.columns
                ]
                dataframe = dataframe[df_order]
                attributes_map_word_to_header = {key: value["abbreviation"] for key, value in self.TASKS_METADATA.items()}
                attributes_map_word_to_header.update(extra_attributes_map_word_to_header)
                attributes_map_word_to_header.update(visible_metrics_map_word_to_header)
                dataframe = dataframe.rename(
                    columns=attributes_map_word_to_header
                )
                return dataframe

    def start_tournament(self, new_submission_id, new_model_file):
        with self.var_lock.ro:
            new_tournament = copy.deepcopy(self.tournament_results)
            new_tournament[new_submission_id] = {}
            new_tournament[new_submission_id][new_submission_id] = {
                task: False for task in self.TASKS_METADATA.keys()
            }
            
            rest_of_competitors = list(self.submission_ids - {new_submission_id})  # without self
            num_of_competitors = len(rest_of_competitors)
            
            result_url = {}
            result_inverse_url = {}
            
            while rest_of_competitors:
                next_competitors = []
                while rest_of_competitors:
                    if len(next_competitors) < 5:  # 5*2==10 tasks
                        next_competitors.append(rest_of_competitors.pop())
                    else:
                        break
                
                for competitor_id in next_competitors:
                    result_url[competitor_id] = check_significance_send_task(new_model_file, self.submission_id_to_file[competitor_id])
                    result_inverse_url[competitor_id] = check_significance_send_task(self.submission_id_to_file[competitor_id], new_model_file)
                
                while next_competitors:
                    competitor_id = next_competitors.pop(0)
                    result = check_significance_wait_for_result(result_url.pop(competitor_id))
                    result_inverse = check_significance_wait_for_result(result_inverse_url.pop(competitor_id))
                    
                    if rest_of_competitors:
                        new_competitor_id = rest_of_competitors.pop()
                        next_competitors.append(new_competitor_id)
                        result_url[new_competitor_id] = check_significance_send_task(new_model_file, self.submission_id_to_file[new_competitor_id])
                        result_inverse_url[new_competitor_id] = check_significance_send_task(self.submission_id_to_file[new_competitor_id], new_model_file)
                    
                    new_tournament[new_submission_id][competitor_id] = result
                    new_tournament[competitor_id][new_submission_id] = result_inverse
                    
                    num_of_competitors_done = num_of_competitors - len(next_competitors) - len(rest_of_competitors)
                    gr.Info(f"Tournament: {num_of_competitors_done}/{num_of_competitors} = {(num_of_competitors_done) * 100 // num_of_competitors}% done")
        
        return new_tournament

    @staticmethod
    def abbreviate(s, max_length, dots_place="center"):
        if len(s) <= max_length:
            return s
        else:
            if max_length <= 1:
                return "…"
            elif dots_place == "begin":
                return "…" + s[-max_length + 1:].lstrip()
            elif dots_place == "center" and max_length >= 3:
                max_length_begin = max_length // 2
                max_length_end = max_length - max_length_begin - 1
                return s[:max_length_begin].rstrip() + "…" + s[-max_length_end:].lstrip()
            else:  # dots_place == "end"
                return s[:max_length - 1].rstrip() + "…"

    @staticmethod
    def create_submission_id(metadata):
        # Délka ID musí být omezena, protože se používá v názvu souboru
        submission_id = "_".join([metadata[key][:7] for key in (
            "team_name",
            "model_name",
            "model_predictions_sha256",
            "model_results_sha256",
        )])
        submission_id = submission_id.replace("/", "_").replace("\n", "_").strip()
        return submission_id

    @staticmethod
    def get_sha256_hexdigest(obj):
        data = json.dumps(
            obj,
            separators=(',', ':'),
            sort_keys=True,
            ensure_ascii=True,
        ).encode()
        result = hashlib.sha256(data).hexdigest()
        return result
    
    PreSubmit = namedtuple('PreSubmit', 'tournament_results, submission_id, file')
    
    def prepare_model_for_submission(self, file, metadata) -> PreSubmit:
        with open(file, "r") as f:
            data = json.load(f)
        
        data["metadata"] = metadata
        
        metadata["model_predictions_sha256"] = self.get_sha256_hexdigest(data["predictions"])
        metadata["model_results_sha256"] = self.get_sha256_hexdigest(data["results"])
        
        submission_id = self.create_submission_id(metadata)
        metadata["submission_id"] = submission_id
        
        metadata["submission_timestamp"] = time.time()  # timestamp
        
        with open(file, "w") as f:
            json.dump(data, f, separators=(',', ':'))  # compact JSON
        
        while True:
            with self.pre_submit_lock:
                if self.pre_submit == None:
                    gr.Info('Running tournament...', duration=15)
                    self.update_leaderboard()
                    tournament_results = self.start_tournament(submission_id, file)
                    self.pre_submit = self.PreSubmit(tournament_results, submission_id, file)
                    break
            gr.Info("Waiting in queue...", duration=5)
            time.sleep(10)
        
        return self.pre_submit

    def save_pre_submit(self):
        with self.pre_submit_lock:
            if self.pre_submit:
                tournament_results, submission_id, file = self.pre_submit
                
                self._upload_submission(submission_id, file)
                self._upload_tournament_results(tournament_results)
                
                self.pre_submit = None
                self.update_leaderboard()

    def _upload_submission(self, submission_id, file):
        api.upload_file(
            path_or_fileobj=file,
            path_in_repo=f"data/{submission_id}.json",
            repo_id=self.SERVER_ADDRESS,
            repo_type=self.REPO_TYPE,
            token=HF_TOKEN,
        )

    def _upload_tournament_results(self, tournament_results):
        # Temporary save tournament results
        with self.results_dataset_local_snapshot_lock.rw:
            tournament_results_path = os.path.join(self.results_dataset_local_snapshot, "tournament.json")
            with open(tournament_results_path, "w") as f:
                json.dump(tournament_results, f, sort_keys=True, indent=2)  # readable JSON

        api.upload_file(
            path_or_fileobj=tournament_results_path,
            path_in_repo="tournament.json",
            repo_id=self.SERVER_ADDRESS,
            repo_type=self.REPO_TYPE,
            token=HF_TOKEN,
        )

    def get_model_detail(self, submission_id):
        with self.var_lock.ro:
            if submission_id not in self.submission_id_to_data:
                raise gr.Error(f"Submission [{submission_id}] not found")
            else:
                data = self.submission_id_to_data[submission_id]
                return data["metadata"]