src/leaderboard/read_evals.py

import glob
import json
import math
import os
from dataclasses import dataclass
from typing import List
import traceback

import dateutil
import numpy as np

from huggingface_hub import ModelCard

from src.display.formatting import make_clickable_model
from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, Language, WeightType, ORIGINAL_TASKS
from src.envs import GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS, SHOW_INCOMPLETE_EVALS

@dataclass
class EvalResult:
    # Also see src.display.utils.AutoEvalColumn for what will be displayed.
    eval_name: str # org_model_precision (uid)
    full_model: str # org/model (path on hub)
    org: str 
    model: str
    revision: str # commit hash, "" if main
    results: dict
    precision: Precision = Precision.Unknown
    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
    weight_type: WeightType = WeightType.Original # Original or Adapter
    main_language: Language = Language.Unknown
    architecture: str = "Unknown" # From config file
    license: str = "?"
    likes: int = 0
    num_params: int = 0
    date: str = "" # submission date of request file
    still_on_hub: bool = True
    is_merge: bool = False
    flagged: bool = False
    status: str = "FINISHED"
    tags: list = None
    json_filename: str = None
    eval_time: float = 0.0
    original_benchmark_average: float = None
    hidden: bool = False # Do not show on the leaderboard

    @classmethod
    def init_from_json_file(self, json_filepath, is_original=False):
        """Inits the result from the specific model result file"""
        with open(json_filepath) as fp:
            data = json.load(fp)

        json_filename = os.path.basename(json_filepath)

        # We manage the legacy config format
        config = data.get("config_general")

        # Precision
        precision = Precision.from_str(config.get("model_dtype"))

        # Get model and org
        org_and_model = config.get("model_name")
        org_and_model = org_and_model.split("/", 1)

        if len(org_and_model) == 1:
            org = None
            model = org_and_model[0]
            result_key = f"{model}_{precision.value.name}"
        else:
            org = org_and_model[0]
            model = org_and_model[1]
            result_key = f"{org}_{model}_{precision.value.name}"
        full_model = "/".join(org_and_model)

        # Extract results available in this file (some results are split in several files)
        results = {}
        tasks = [(task.value.benchmark, task.value.metric) for task in Tasks]
        if is_original:
            tasks = ORIGINAL_TASKS
        for task in tasks:
            benchmark, metric = task
            
            # We skip old mmlu entries
            wrong_mmlu_version = False
            if benchmark == "hendrycksTest":
                for mmlu_k in ["harness|hendrycksTest-abstract_algebra|5", "hendrycksTest-abstract_algebra"]:
                    if mmlu_k in data["versions"] and data["versions"][mmlu_k] == 0:
                        wrong_mmlu_version = True

            if wrong_mmlu_version:
                continue

            # Some truthfulQA values are NaNs
            if benchmark == "truthfulqa:mc" and "harness|truthfulqa:mc|0" in data["results"]:
                if math.isnan(float(data["results"]["harness|truthfulqa:mc|0"][metric])):
                    results[benchmark] = 0.0
                    continue

            def get_metric(v):
                res = v.get(metric, None)
                if res is None:
                    res = v.get(metric + ',all', None)
                if res is None:
                    res = v.get(metric + ',None', None)
                if res is None:
                    res = v.get('main_score', None)
                return res
                
            # We average all scores of a given metric (mostly for mmlu)
            accs = np.array([get_metric(v) for k, v in data["results"].items() if benchmark in k])
            if accs.size == 0 or any([acc is None for acc in accs]):
                continue

            mean_acc = np.mean(accs) * 100.0
            results[benchmark] = mean_acc

        return self(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            results=results,
            precision=precision,  
            revision= config.get("model_sha", ""),
            json_filename=json_filename,
            eval_time=config.get("total_evaluation_time_seconds", 0.0)
        )

    def update_with_request_file(self, requests_path):
        """Finds the relevant request file for the current model and updates info with it"""
        request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)

        try:
            with open(request_file, "r") as f:
                request = json.load(f)
            self.model_type = ModelType.from_str(request.get("model_type", "Unknown"))
            self.weight_type = WeightType[request.get("weight_type", "Original")]
            self.num_params = request.get("params", 0)
            self.date = request.get("submitted_time", "")
            self.architecture = request.get("architectures", "Unknown")
            self.status = request.get("status", "FAILED")
            self.hidden = request.get("hidden", False)
            self.main_language = request.get("main_language", "?")
        except Exception as e:
            self.status = "FAILED"
            print(f"Could not find request file for {self.org}/{self.model}")

    def update_with_dynamic_file_dict(self, file_dict):
        self.license = file_dict.get("license", "?")
        self.likes = file_dict.get("likes", 0)
        self.still_on_hub = file_dict["still_on_hub"]
        self.flagged = any("flagged" in tag for tag in file_dict["tags"])
        self.tags = file_dict["tags"]
        if 'original_llm_scores' in file_dict:
            if len(file_dict['original_llm_scores']) > 0:
                if self.precision.value.name in file_dict['original_llm_scores']:
                    self.original_benchmark_average = file_dict['original_llm_scores'][self.precision.value.name]
                else:
                    self.original_benchmark_average = max(list(file_dict['original_llm_scores'].values()))
        
    def to_dict(self):
        """Converts the Eval Result to a dict compatible with our dataframe display"""
        average = []
        npm = []
        for task in Tasks:
            if task.value.benchmark not in self.results:
                continue
            res = self.results[task.value.benchmark]
            if res is None or np.isnan(res) or not (isinstance(res, float) or isinstance(res, int)):
                continue
            average.append(res)
            npm.append((res-task.value.baseline)*100.0 / (100.0-task.value.baseline))
        average = round(sum(average)/len(average), 2)
        npm = round(sum(npm)/len(npm), 2)

        data_dict = {
            "eval_name": self.eval_name,  # not a column, just a save name,
            AutoEvalColumn.precision.name: self.precision.value.name,
            AutoEvalColumn.model_type.name: self.model_type.value.name,
            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
            AutoEvalColumn.weight_type.name: self.weight_type.value.name,
            AutoEvalColumn.architecture.name: self.architecture,
            AutoEvalColumn.model.name: make_clickable_model(self.full_model, self.json_filename),
            AutoEvalColumn.dummy.name: self.full_model,
            AutoEvalColumn.revision.name: self.revision,
            AutoEvalColumn.average.name: average,
            AutoEvalColumn.license.name: self.license,
            AutoEvalColumn.likes.name: self.likes,
            AutoEvalColumn.params.name: self.num_params,
            AutoEvalColumn.still_on_hub.name: self.still_on_hub,
            AutoEvalColumn.merged.name: "merge" in self.tags if self.tags else False,
            AutoEvalColumn.moe.name: ("moe" in self.tags if self.tags else False) or "moe" in self.full_model.lower(),
            AutoEvalColumn.flagged.name: self.flagged,
            AutoEvalColumn.eval_time.name: self.eval_time,
            AutoEvalColumn.npm.name: npm,
            AutoEvalColumn.main_language.name: self.main_language
        }

        for task in Tasks:
            if task.value.benchmark in self.results:
                data_dict[task.value.col_name] = self.results[task.value.benchmark]

        if GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS:
            data_dict[AutoEvalColumn.original_benchmark_average.name] = self.original_benchmark_average

        return data_dict


def get_request_file_for_model(requests_path, model_name, precision):
    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
    request_files = os.path.join(
        requests_path,
        f"{model_name}_eval_request_*.json",
    )
    request_files = glob.glob(request_files)

    # Select correct request file (precision)
    request_file = ""
    request_files = sorted(request_files, reverse=True)
    for tmp_request_file in request_files:
        with open(tmp_request_file, "r") as f:
            req_content = json.load(f)
            if (
                req_content["status"] in ["FINISHED", "PENDING_NEW_EVAL" if SHOW_INCOMPLETE_EVALS else "FINISHED"]
                and req_content["precision"] == precision.split(".")[-1]
            ):
                request_file = tmp_request_file
    return request_file


def get_raw_eval_results(results_path: str, requests_path: str, dynamic_path: str) -> list[EvalResult]:
    """From the path of the results folder root, extract all needed info for results"""
    model_result_filepaths = []

    for root, _, files in os.walk(results_path):
        # We should only have json files in model results
        if len(files) == 0 or any([not f.endswith(".json") for f in files]):
            continue

        # Sort the files by date
        try:
            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
        except dateutil.parser._parser.ParserError:
            files = [files[-1]]

        for file in files:
            model_result_filepaths.append(os.path.join(root, file))

    with open(dynamic_path) as f:
        dynamic_data = json.load(f)

    eval_results = {}
    for model_result_filepath in model_result_filepaths:
        # Creation of result
        eval_result = EvalResult.init_from_json_file(model_result_filepath)
        eval_result.update_with_request_file(requests_path)
        if eval_result.full_model in dynamic_data:
            eval_result.update_with_dynamic_file_dict(dynamic_data[eval_result.full_model])

        # Store results of same eval together
        eval_name = eval_result.eval_name
        if eval_name in eval_results.keys():
            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
            eval_results[eval_name].json_filename = eval_result.json_filename
        else:
            eval_results[eval_name] = eval_result

    results = []
    for v in eval_results.values():
        try:
            if v.status in ["FINISHED", "PENDING_NEW_EVAL" if SHOW_INCOMPLETE_EVALS else "FINISHED"] and not v.hidden:
                v.to_dict() # we test if the dict version is complete
                results.append(v)
        except KeyError as e:  # not all eval values present
            continue

    return results