src/leaderboard/read_evals.py

import glob
import json
import os
from dataclasses import dataclass

import numpy as np

from src.display.formatting import make_clickable_model
from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
from src.submission.check_validity import is_model_on_hub


@dataclass
class EvalResult:
    """Represents one full evaluation. Built from a combination of the result and request file for a given run.
    """
    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
    average_accuracy: float
    precision: Precision = Precision.Unknown
    model_type: ModelType = ModelType.Unknown  # Pretrained, fine tuned, ...
    weight_type: WeightType = WeightType.Original  # Original or Adapter
    architecture: str = "Unknown"
    license: str = "?"
    likes: int = 0
    num_params: int = 0
    date: str = ""  # submission date of request file
    still_on_hub: bool = False

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

        config = data.get("config", {})

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

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

        results_data = data.get("results", {})

        # Extract per-subject accuracies
        per_subject_results = {}
        for task in Tasks:
            subject = task.value.benchmark
            accuracy = results_data.get(subject, None)
            if accuracy is not None:
                per_subject_results[subject] = accuracy

        average_accuracy = results_data.get('average', None)

        # Set other fields from config
        model_type = ModelType.from_str(config.get("model_type", ""))
        weight_type = WeightType[config.get("weight_type", "Original")]
        license = config.get("license", "?")
        likes = config.get("likes", 0)
        num_params = config.get("params", 0)
        date = config.get("submitted_time", "")
        still_on_hub = config.get("still_on_hub", True)
        architecture = config.get("architecture", "Unknown")

        # Create EvalResult instance
        return cls(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            results=per_subject_results,
            average_accuracy=average_accuracy,
            precision=precision,
            revision=config.get("model_sha", ""),
            still_on_hub=still_on_hub,
            architecture=architecture,
            model_type=model_type,
            weight_type=weight_type,
            license=license,
            likes=likes,
            num_params=num_params,
            date=date,
        )

    def to_dict(self):
        """Converts the Eval Result to a dict compatible with our dataframe display"""
        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),
            AutoEvalColumn.revision.name: self.revision,
            AutoEvalColumn.average.name: self.average_accuracy,
            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,
        }

        for task in Tasks:
            subject = task.value.benchmark
            data_dict[task.value.col_name] = self.results.get(subject, None)

        return data_dict


def get_raw_eval_results(results_path: str, requests_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
        for file in files:
            if file.endswith(".json"):
                model_result_filepaths.append(os.path.join(root, file))

    eval_results = {}
    for model_result_filepath in model_result_filepaths:
        # Creation of result
        eval_result = EvalResult.init_from_json_file(model_result_filepath)
        # Store results
        eval_name = eval_result.eval_name
        eval_results[eval_name] = eval_result

    results = []
    for v in eval_results.values():
        try:
            v.to_dict()  # we test if the dict version is complete
            results.append(v)
        except KeyError:  # not all eval values present
            continue

    return results