src/leaderboard/read_evals.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# flake8: noqa E501

import glob
import json
import os
from dataclasses import dataclass

import dateutil
import numpy as np

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


@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)
    model_name: 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
    architecture: str = "Unknown"
    license: str = "Unknown"
    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)

        if 'human_eval_solidity_pass_1' not in data['results']:
            data['results']['human_eval_solidity_pass_1'] = {'score': 0}
            
        if 'human_eval_solidity_pass_3' not in data['results']:
            data['results']['human_eval_solidity_pass_3'] = {'score': 0}

        org, model = get_org_and_model_names_from_filepath(json_filepath)
        config = data.get("config")

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

        result_key = f"{org}_{model}_{precision.value.name}"
        model_name = get_model_name_from_filepath(json_filepath)

        still_on_hub, _, model_config = is_model_on_hub(
            model_name,
            config.get("model_sha", "main"),
            trust_remote_code=True,
            test_tokenizer=False,
        )
        architecture = "Unknown"
        if model_config is not None:
            architectures = getattr(model_config, "architectures", None)
            if architectures:
                architecture = ";".join(architectures)

        # Extract results available in this file
        # (some results are split in several files)
        results = {}
        for task in Tasks:
            task = task.value

            # We average all scores of a given metric
            # (not all metrics are present in all files)
            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])
            if accs.size == 0 or any([acc is None for acc in accs]):
                continue

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

        return cls(
            eval_name=result_key,
            model_name=model_name,
            org=org,
            model=model,
            results=results,
            precision=precision,
            revision=config.get("model_sha", ""),
            still_on_hub=still_on_hub,
            architecture=architecture
        )

    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.model_name,
            self.revision,
            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", ""))
            self.weight_type = WeightType[request.get("weight_type", "Original")]
            self.license = request.get("license", "Unknown")
            self.likes = request.get("likes", 0)
            self.num_params = request.get("params", 0)
            self.date = request.get("submitted_time", "")
        except Exception as error:
            print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")
            print(f"Error: {error}")

    def to_dict(self):
        """Converts the Eval Result to a dict compatible with our dataframe display"""
        # average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
        scores = {
            'naive_judge': self.results.get('naive_judge', 0),
            'human_eval_solidity_pass_1': self.results.get('human_eval_solidity_pass_1', 0),
            'human_eval_solidity_pass_3': self.results.get('human_eval_solidity_pass_3', 0)
        }
    
        solbench = 0
        non_zero_scores = {k: v for k, v in scores.items() if v != 0}
        if non_zero_scores:
            weights = {
                'naive_judge': 0.3,
                'human_eval_solidity_pass_1': 0.5,
                'human_eval_solidity_pass_3': 0.2
            }
            total_weight = sum(weights[k] for k in non_zero_scores)
            solbench = sum(scores[k] * weights[k] / total_weight for k in non_zero_scores)

        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.model_name),
            AutoEvalColumn.revision.name: self.revision,
            # AutoEvalColumn.average.name: average,
            AutoEvalColumn.solbench.name: solbench,
            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:
            data_dict[task.value.col_name] = self.results[task.value.benchmark]

        return data_dict


def get_request_file_for_model(
        requests_path: str,
        model_name: str,
        revision: str,
        precision: str,
    ):
    request_hash = get_request_hash(model_name, revision, precision)
    filepath = os.path.join(requests_path, model_name, '{}.json'.format(request_hash))
    print(f'Loading {filepath}...')
    filepath = glob.glob(filepath)[0]
    return filepath


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
        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))

    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)

        # 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})
        else:
            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