backend/app/leaderboard/read_evals.py

import glob
import json
import os
from dataclasses import dataclass
from functools import lru_cache

import numpy as np

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


# Add caching for hub checks to avoid repeated network calls
@lru_cache(maxsize=256)
def cached_is_model_on_hub(full_model, revision):
    """Cached version of is_model_on_hub to avoid repeated network calls"""
    return is_model_on_hub(full_model, revision, trust_remote_code=True, test_tokenizer=False)


@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
    model_type: ModelType = ModelType.Unknown  # Pretrained, fine tuned, ...
    architecture: str = "Unknown"
    license: str = "?"
    likes: int = 0
    num_params: int = 0
    date: str = ""  # submission date of request file
    still_on_hub: bool = False
    reasoning: bool = False  # Whether reasoning is enabled for this model
    note: str = ""  # Extra information about the model (e.g., thinking budget, warnings)

    @classmethod
    def init_from_new_format_json_file(self, json_filepath):
        """Inits the result from the new format model result file"""
        with open(json_filepath) as fp:
            data = json.load(fp)

        results = data.get("results")

        full_model = data.get("config_general", {}).get("model_name", "").strip()
        result_key = full_model.replace("/", "_")

        org, model = full_model.split("/", 1) if "/" in full_model else ("", full_model)

        still_on_hub, _, model_config = cached_is_model_on_hub(full_model, "main")

        architecture = "?"
        if model_config is not None:
            architectures = getattr(model_config, "architectures", None)
            if architectures:
                architecture = ";".join(architectures)

        # Extract results available in this file
        score_results = {}
        for task in Tasks:
            task = task.value
            benchmark_id = task.benchmark
            metric = task.metric

            scores = [
                results[key][metric]
                for key in results
                if "|" in key and benchmark_id.startswith(key.split("|")[1].removeprefix("icelandic_evals:"))
            ]
            if len(scores) == 0:
                continue

            mean_acc = np.mean(scores) * 100.0
            score_results[benchmark_id] = mean_acc

        return self(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            results=score_results,
            revision="",
            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.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", ""))
            self.license = request.get("license", "?")
            self.likes = request.get("likes", 0)
            self.num_params = request.get("params", 0)
            self.date = request.get("submitted_time", "")
            self.reasoning = request.get("reasoning", False) or request.get("gen_kwargs", {}).get(
                "reasoning_effort", None
            )
            self.note = request.get("note", "")  # Default to empty string if missing
        except FileNotFoundError:
            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.model_type.name: self.model_type.value.name,
            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
            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,
            AutoEvalColumn.reasoning.name: self.reasoning,
            AutoEvalColumn.note.name: self.note,
        }

        for task in Tasks:
            if task.value.benchmark in self.results.keys():
                data_dict[task.value.col_name] = self.results[task.value.benchmark]
            else:
                data_dict[task.value.col_name] = None

        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)
    if len(request_files) == 1:
        return request_files[0]

    # 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["precision"] == precision.split(".")[-1] or req_content["precision"] is None:
                request_file = tmp_request_file
    return request_file


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 = []

    # Collect all JSON files first
    for root, _, files in os.walk(results_path):
        # We should only have json files in model results
        json_files = [f for f in files if f.endswith(".json")]
        if len(json_files) == 0:
            continue

        # Sort JSON files by date (newer later)
        try:
            json_files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
        except (ValueError, IndexError):
            # If sorting fails, just use the files as-is or take the last one
            json_files = [json_files[-1]] if json_files else []

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

    eval_results = {}
    for model_result_filepath in model_result_filepaths:
        try:
            # Creation of result
            eval_result = EvalResult.init_from_new_format_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:
                # Update with newer scores
                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
        except Exception as e:
            # Log error but continue processing other files
            print(f"Error processing {model_result_filepath}: {e}")
            continue

    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