src/leaderboard/read_evals.py

import glob
import json
import os
from collections import defaultdict
from dataclasses import dataclass
from pathlib import Path

import numpy as np

from src.display.formatting import make_clickable_model
from src.display.utils import AutoEvalColumn, ModelTraining, Tasks, Precision, WeightType, MalteseTraining
from src.envs import TOKEN, API
from src.submission.check_validity import is_model_on_hub, get_model_size


@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
    precision: Precision = Precision.Unknown
    n_shot: int = 0
    prompt_version: str = "1.0_english"
    seed: int = 0
    model_training: ModelTraining = ModelTraining.NK # Pretrained, fine tuned, ...
    maltese_training: MalteseTraining = MalteseTraining.NK # none, pre-training, ...
    language_count: int = None
    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_files(self, seed_directory):
        """Inits the result from the specific model result file"""
        with open(list(seed_directory.values())[0][0]) as fp:
            data = json.load(fp)

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

        n_shot = config.get("n_shot")

        prompt_version = config.get("prompt_version")

        seed = config.get("seed")

        model_training = ModelTraining.from_str(config.get("model_training"))

        maltese_training = MalteseTraining.from_str(config.get("maltese_training"))

        language_count = config.get("language_count")

        model_size = config.get("model_num_parameters")

        # Get model and org
        org_and_model = config.get("model", None)
        org_and_model = org_and_model.split("/", 1)
        full_model = "/".join(org_and_model)

        revision = config.get("model_sha", config.get("model_revision", "main"))

        model_args = config.get("model_args")
        model_args["revision"] = revision
        model_args["trust_remote_code"] = True
        model_args["cache_dir"] = None
        base_model = None
        if "pretrained" in model_args:
            base_model = model_args.pop("pretrained")
        still_on_hub, _, model_config = is_model_on_hub(
            base_model or full_model, model_args, test_tokenizer=False, token=TOKEN,
        )
        architecture = "?"
        if model_config is not None:
            architectures = getattr(model_config, "architectures", None)
            if architectures:
                architecture = ";".join(architectures)
        license = "?"
        likes = 0
        if still_on_hub:
            try:
                model_info = API.model_info(repo_id=full_model, revision=revision, token=TOKEN)
                if not model_size:
                    model_size = get_model_size(model_info=model_info, precision=precision)
                license = model_info.cardData.get("license")
                likes = model_info.likes
            except Exception:
                pass

        # Extract results available in this file (some results are split in several files)
        results = defaultdict(dict)
        for seed, file_paths in seed_directory.items():
            for file_path in file_paths:
                with open(file_path) as file:
                    data = json.load(file)["results"]

                for task in Tasks:
                    task = task.value
                    if task.benchmark not in data or task.metric not in data[task.benchmark]:
                        continue
                    score = data[task.benchmark][task.metric]
                    if task.metric in ("acc", "f1", "loglikelihood", "rouge"):
                        score *= 100
                    results[task.benchmark + "_" + task.metric][seed] = score

        results = {task: np.mean(list(seed_results.values())) for task, seed_results in results.items()}

        if len(org_and_model) == 1:
            org = None
            model = org_and_model[0]
        else:
            org = org_and_model[0]
            model = org_and_model[1]
        result_key = f"{'_'.join(org_and_model)}_{revision}_{precision.value.name}_{n_shot}_{prompt_version}_{seed}"

        return self(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            results=results,
            model_training=model_training,
            maltese_training=maltese_training,
            language_count=language_count or "?",
            precision=precision,
            revision=revision,
            n_shot=n_shot,
            prompt_version=prompt_version,
            seed=seed,
            still_on_hub=still_on_hub,
            architecture=architecture,
            likes=likes or "?",
            num_params=model_size and round(model_size / 1e9, 3),
            license=license,
        )

    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_training = ModelTraining.from_str(request.get("model_training", ""))
            self.weight_type = WeightType[request.get("weight_type", "Original")]
            self.license = request.get("license", "?")
            self.likes = request.get("likes", 0)
            self.num_params = request.get("params", 0)
            self.date = request.get("submitted_time", "")
        except Exception:
            print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")

    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)
        data_dict = {
            "eval_name": self.eval_name,  # not a column, just a save name,
            AutoEvalColumn.precision.name: self.precision.value.name,
            AutoEvalColumn.n_shot.name: self.n_shot,
            AutoEvalColumn.prompt_version.name: self.prompt_version,
            AutoEvalColumn.model_training.name: self.model_training.value.name,
            AutoEvalColumn.maltese_training.name: self.maltese_training.value.name,
            AutoEvalColumn.model_symbol.name: self.model_training.value.symbol + "/" + self.maltese_training.value.symbol,
            AutoEvalColumn.language_count.name: self.language_count,
            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: 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,
        }

        results_by_task_type = defaultdict(list)
        for task in Tasks:
            result = self.results.get(task.value.benchmark + "_" + task.value.metric)
            data_dict[task.value.col_name] = result
            if task.value.is_primary_metric and not (task.value.zero_shot_only and self.n_shot > 0):
                results_by_task_type[task.value.task_type].append(result)
        results_averages = []
        for task_type, task_type_results in results_by_task_type.items():
            average = sum([score for score in task_type_results if score is not None]) / len(task_type_results)
            data_dict[getattr(AutoEvalColumn, task_type.value.name).name] = average
            results_averages.append(average)
        data_dict[AutoEvalColumn.average.name] = np.mean(results_averages) if len(results_averages) > 1 else results_averages[0]

        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"]
                and req_content["precision"] == precision.split(".")[-1]
            ):
                request_file = tmp_request_file
    return request_file


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

    for directory_path in Path(results_path).rglob("*/*/"):
        for file_path in directory_path.rglob("*-seed/results_*.json"):
            seed = file_path.parent.name.removesuffix("-seed")
            model_result_filepaths[directory_path.relative_to(results_path)][seed].append(file_path)

    eval_results = {}
    for model_result_filepath in model_result_filepaths.values():
        # Creation of result
        eval_result = EvalResult.init_from_json_files(model_result_filepath)

        # 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