src/display/utils.py

from dataclasses import dataclass, make_dataclass, field
from enum import Enum
from typing import List
import pandas as pd
import os
import json
from copy import deepcopy
from yaml import safe_load
from src.envs import GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS, TASK_CONFIG

def fields(raw_class):
    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]


@dataclass
class Task:
    benchmark: str
    metric: str
    col_name: str
    baseline: float = 0.0
    human_baseline: float = None
    expert_human_baseline: float = None
    few_shot: int = None
    limit: int = None
    task_list: List[str] = None
    link: str = None
    description: str = None
    sources: List[str] = None
    baseline_sources: List[str] = None
    citation: str = None

Tasks = Enum('Tasks', {k: Task(**v) for k, v in TASK_CONFIG['tasks'].items()})

# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    dummy: bool = False

auto_eval_column_dict = []
# Init
auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
#Scores
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
# Model information
auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
auto_eval_column_dict.append(["merged", ColumnContent, ColumnContent("Merged", "bool", False)])
auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, hidden=True)])
auto_eval_column_dict.append(["model_sha", ColumnContent, ColumnContent("Model sha", "str", False, False)])
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Revision", "str", False, False)])
auto_eval_column_dict.append(["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
auto_eval_column_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
auto_eval_column_dict.append(["eval_time", ColumnContent, ColumnContent("Evaluation Time (s)", "number", False)])
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("Model Name", "str", False, dummy=True)])
if GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS:
    auto_eval_column_dict.append(["original_benchmark_average", ColumnContent, ColumnContent("🤗 Leaderboard Average", "number", False)])
auto_eval_column_dict.append(["npm", ColumnContent, ColumnContent("NPM (Average) ⬆️", "number", False)])
auto_eval_column_dict.append(["main_language", ColumnContent, ColumnContent("Main Language", "str", False)])

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn",
                                [(name, ae_class, field(default_factory=lambda: value)) for name, ae_class, value in auto_eval_column_dict], 
                                frozen=True)
for name, ae_class, value in auto_eval_column_dict:
    setattr(AutoEvalColumn, name, value)



@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model = ColumnContent("model", "markdown", True)
    revision = ColumnContent("revision", "str", True)
    private = ColumnContent("private", "bool", True)
    precision = ColumnContent("precision", "str", True)
    weight_type = ColumnContent("weight_type", "str", "Original")
    status = ColumnContent("status", "str", True)


baseline_row = {
    AutoEvalColumn.model.name: "<p>Baseline</p>",
    AutoEvalColumn.revision.name: "N/A",
    AutoEvalColumn.model_sha.name: "N/A",
    AutoEvalColumn.precision.name: "?",
    AutoEvalColumn.merged.name: False,
    #AutoEvalColumn.average.name: 31.0,
    AutoEvalColumn.dummy.name: "baseline",
    AutoEvalColumn.model_type.name: "",
    AutoEvalColumn.flagged.name: False,
    AutoEvalColumn.model_type_symbol.name: "?",
    AutoEvalColumn.architecture.name: None,
    AutoEvalColumn.weight_type.name: None,
    AutoEvalColumn.params.name: 0,
    AutoEvalColumn.likes.name: 0,
    AutoEvalColumn.license.name: "",
    AutoEvalColumn.still_on_hub.name: False,
    AutoEvalColumn.moe.name: False,
    AutoEvalColumn.eval_time.name: 0.0,
    AutoEvalColumn.main_language.name: "?",
    'hf_path': None,
}

baseline_list = []
npm = []
for task in Tasks:
    baseline_row[task.value.col_name] = task.value.baseline
    res = task.value.baseline
    if res is not None and (isinstance(res, float) or isinstance(res, int)):
        baseline_list.append(res)
        npm.append((res - task.value.baseline) / (100 - task.value.baseline))
baseline_row[AutoEvalColumn.average.name] = round(sum(baseline_list) / len(baseline_list), 2)
baseline_row[AutoEvalColumn.npm.name] = round(sum(npm) / len(npm), 2)

if GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS:
    baseline_row["🤗 Leaderboard Average"] = None

# Average ⬆️ human baseline is 0.897 (source: averaging human baselines below)
# ARC human baseline is 0.80 (source: https://lab42.global/arc/)
# HellaSwag human baseline is 0.95 (source: https://deepgram.com/learn/hellaswag-llm-benchmark-guide)
# MMLU human baseline is 0.898 (source: https://openreview.net/forum?id=d7KBjmI3GmQ)
# TruthfulQA human baseline is 0.94(source: https://arxiv.org/pdf/2109.07958.pdf)
# Winogrande: https://leaderboard.allenai.org/winogrande/submissions/public
# GSM8K: paper
# Define the human baselines
human_baseline_row = {
    AutoEvalColumn.model.name: "<p>Human performance</p>",
    AutoEvalColumn.revision.name: "N/A",
    AutoEvalColumn.model_sha.name: "N/A",
    AutoEvalColumn.precision.name: "?",
    #AutoEvalColumn.average.name: 92.75,
    AutoEvalColumn.merged.name: False,
    AutoEvalColumn.dummy.name: "human_baseline",
    AutoEvalColumn.model_type.name: "",
    AutoEvalColumn.flagged.name: False,
    AutoEvalColumn.model_type_symbol.name: "?",
    AutoEvalColumn.architecture.name: None,
    AutoEvalColumn.weight_type.name: None,
    AutoEvalColumn.params.name: 0,
    AutoEvalColumn.likes.name: 0,
    AutoEvalColumn.license.name: "",
    AutoEvalColumn.still_on_hub.name: False,
    AutoEvalColumn.moe.name: False,
    AutoEvalColumn.eval_time.name: 0.0,
    AutoEvalColumn.main_language.name: "?",
    'hf_path': None,
}

baseline_list = []
npm = []
for task in Tasks:
    human_baseline_row[task.value.col_name] = task.value.human_baseline
    res = task.value.human_baseline
    if res is None or not (isinstance(res, float) or isinstance(res, int)):
        res = 95.0
    baseline_list.append(res)
    npm.append((res - task.value.baseline) / (100 - task.value.baseline))
human_baseline_row[AutoEvalColumn.average.name] = round(sum(baseline_list) / len(baseline_list), 2)
human_baseline_row[AutoEvalColumn.npm.name] = round(sum(npm) / len(npm), 2)
if GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS:
    human_baseline_row["🤗 Leaderboard Average"] = None

@dataclass
class ModelDetails:
    name: str
    symbol: str = "" # emoji, only for the model type


class ModelType(Enum):
    PT = ModelDetails(name="pretrained", symbol="🟢")
    LA = ModelDetails(name="language adapted (FP, FT, ...)", symbol="🆎")
    FT = ModelDetails(name="fine-tuned/fp on domain-specific datasets", symbol="🔶")
    chat = ModelDetails(name="chat (RLHF, DPO, IFT, ...)", symbol="💬")
    merges = ModelDetails(name="base merges and moerges", symbol="🤝")
    proprietary = ModelDetails(name="proprietary (closed)", symbol="🔒")
    Unknown = ModelDetails(name="", symbol="?")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        if "fine-tuned" in type or "🔶" in type:
            return ModelType.FT
        if "language" in type or "🆎" in type:
            return ModelType.LA
        if "pretrained" in type or "🟢" in type:
            return ModelType.PT
        if any([k in type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]):
            return ModelType.chat
        if "merge" in type or "🤝" in type:
            return ModelType.merges
        if "proprietary" in type or "🔒" in type:
            return ModelType.proprietary
        return ModelType.Unknown

class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")

class Precision(Enum):
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")
    qt_8bit = ModelDetails("8bit")
    qt_4bit = ModelDetails("4bit")
    qt_GPTQ = ModelDetails("GPTQ")
    Unknown = ModelDetails("?")

    def from_str(precision):
        if precision in ["torch.float16", "float16"]:
            return Precision.float16
        if precision in ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["8bit"]:
            return Precision.qt_8bit
        if precision in ["4bit"]:
            return Precision.qt_4bit
        if precision in ["GPTQ", "None"]:
            return Precision.qt_GPTQ
        return Precision.Unknown
        
class Language(Enum):
    Portuguese = ModelDetails("Portuguese")
    English = ModelDetails("English")
    Chinese = ModelDetails("Chinese")
    Spanish = ModelDetails("Spanish")
    Other = ModelDetails("Other")
    Unknown = ModelDetails("?")

    def from_str(language):
        language = language.lower().replace('-', '').replace('_', '')
        if language in ["pt", "ptpt", "ptbr", "portuguese"]:
            return Language.Portuguese
        if language in ["en", "enus", "engb", "english"]:
            return Language.English
        if language in ["es", "spanish"]:
            return Language.Spanish
        if language in ["zh", "chinese"]:
            return Language.Chinese
        if language in ["other", "multi", "multilingual"]:
            return Language.Other
        return Language.Unknown

#External models
external_rows = []
external_eval_results = []  # Initialize the list to store EvalResult objects
if os.path.exists('external_models_results.json'):
    with open('external_models_results.json', 'r', encoding='utf8') as f:
        all_models = json.load(f)
    for model_data in all_models:
        #Create external_rows
        model_row = deepcopy(baseline_row)
        model_row[AutoEvalColumn.model.name] = f'<a target="_blank" href="{model_data["link"]}" style="color: var(--text-color); text-decoration: underline;text-decoration-style: dotted;">{model_data["name"]} [{model_data["date"]}]</a>'
        model_row[AutoEvalColumn.dummy.name] = model_data['name']
        for task in Tasks:
            model_row[task.value.col_name] = round(model_data['result_metrics'][task.value.benchmark]*100, 2)
        model_row[AutoEvalColumn.average.name] = round(model_data['result_metrics_average']*100, 2)
        model_row[AutoEvalColumn.npm.name] = round(model_data['result_metrics_npm']*100, 2)
        
        model_type = ModelType.from_str(model_data['model_type'])
        model_row[AutoEvalColumn.model_type.name] = model_type.value.name
        model_row[AutoEvalColumn.model_type_symbol.name] = model_type.value.symbol
        if model_type == ModelType.proprietary:
            model_row[AutoEvalColumn.license.name] = "Proprietary"
        if 'params' in model_data:
            model_row[AutoEvalColumn.params.name] = model_data['params']

        model_row[AutoEvalColumn.main_language.name] = model_data['main_language']
        model_row['hf_path'] = None if 'huggingface.co' not in model_data['link'] else model_data['link'].split('huggingface.co/')[1]
        external_rows.append(model_row)

        #Create external_eval_results
        eval_result = dict(
            eval_name=f"external_{model_data['model']}",
            full_model=model_data['name'],
            org="External",  # External models don't have an org in this context
            model=model_data['name'],
            # Scale results by 100 to match expected format
            results={k: v * 100 for k, v in model_data['result_metrics'].items()},
            model_sha="",  # Not available
            revision="main", # Default
            precision=Precision.Unknown, # Not available
            model_type=model_type, # Already determined above
            weight_type=WeightType.Original, # Assuming original weights
            main_language=model_data['main_language'],
            architecture="Unknown", # Not available
            license="Proprietary" if model_type == ModelType.proprietary else "?",
            likes=0, # Not available
            num_params=model_data.get('params', 0), # Use .get() for safety
            date=model_data['date']+"T00:00:00Z",
            still_on_hub=True, # Not applicable
            is_merge=False, # Not applicable
            flagged=False, # Not applicable
            status="FINISHED",
            tags=None, # Not available
            json_filename='external_models_results.json', # Not applicable
            eval_time=0.0, # Not available
            # Scale average by 100
            original_benchmark_average=None,#model_data.get('result_metrics_average', 0.0) * 100,
            hidden=False, # Default
            num_evals_model_rev=1 # Default
        )
        """
        EvalResult(eval_name='01-ai_Yi-1.5-34B_bfloat16', 
        '   full_model='01-ai/Yi-1.5-34B', 
        '   org='01-ai', 
        '   model='Yi-1.5-34B', 
        '   results={'enem_challenge': 71.51854443666899, 
        '   'bluex': 66.62030598052851, 
        '   'oab_exams': 54.89749430523918, 
        '   'assin2_rte': 89.76911637262349, 
        '   'assin2_sts': 81.48786802023537, 
        '   'faquad_nli': 58.5644163957417, 
        '   'hatebr_offensive': 83.63023241432246, 
        '   'portuguese_hate_speech': 69.62399848962205, 
        '   'tweetsentbr': 72.28749707523902}, 
        '   model_sha='81136a42efdf6f6a63031ac31639a37813fe6e37', 
        '   revision='main', 
        '   precision=<Precision.bfloat16: ModelDetails(name='bfloat16', 
        '   symbol='')>, 
        '   model_type=<ModelType.PT: ModelDetails(name='pretrained', 
        '   symbol='🟢')>, 
        '   weight_type=<WeightType.Original: ModelDetails(name='Original', 
        '   symbol='')>, 
        '   main_language='English', 
        '   architecture='LlamaForCausalLM', 
        '   license='?', 
        '   likes=0, 
        '   num_params=34.39, 
        '   date='2024-05-15T17:40:15Z', 
        '   still_on_hub=True, 
        '   is_merge=False, 
        '   flagged=False, 
        '   status='FINISHED', 
        '   tags=None, 
        '   json_filename='results_2024-05-17T10-36-18.336343.json', 
        '   eval_time=11545.340715408325, 
        '   original_benchmark_average=None, 
        '   hidden=False, 
        '   num_evals_model_rev=1)
        """
        external_eval_results.append(eval_result)


# Column selection
COLS = [c.name for c in fields(AutoEvalColumn)] + ['hf_path']
TYPES = [c.type for c in fields(AutoEvalColumn)]

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

BENCHMARK_COLS = [t.value.col_name for t in Tasks]

NUMERIC_INTERVALS = {
    "?": pd.Interval(-1, 0, closed="right"),
    "~1B": pd.Interval(0, 2, closed="right"),
    "~3B": pd.Interval(2, 5, closed="right"),
    "~8B": pd.Interval(5, 10, closed="right"),
    "~16B": pd.Interval(10, 23, closed="right"),
    "~35B": pd.Interval(23, 50, closed="right"),
    "~70B": pd.Interval(50, 90, closed="right"),
    "100B+": pd.Interval(90, 10000, closed="right"),
}

#Original HF LEaderboard tasks and metrics
ORIGINAL_TASKS = [
    ("arc:challenge", "acc_norm"),
    ("hellaswag", "acc_norm"),
    ("hendrycksTest", "acc"),
    ("truthfulqa:mc", "mc2"),
    ("winogrande", "acc"),
    ("gsm8k", "acc")
]