from dataclasses import dataclass, make_dataclass from enum import Enum import pandas as pd 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 class Tasks(Enum): gpqa = Task("ko_gpqa_diamond_zeroshot", "acc_norm,none", "Ko-GPQA") winogrande = Task("ko_winogrande", "acc,none", "Ko-Winogrande") gsm8k = Task("ko_gsm8k", "exact_match,strict-match", "Ko-GSM8k") eqBench = Task("ko_eqbench", "eqbench,none", "Ko-EQ Bench") instFollow = Task("ko_ifeval", "strict_acc,none", "Ko-IFEval") korNatCka = Task("kornat_common", "acc_norm,none", "KorNAT-CKA") korNatSva = Task("kornat_social", "A-SVA,none", "KorNAT-SVA") harmlessness = Task("kornat_harmless", "acc_norm,none", "Ko-Harmlessness") helpfulness = Task("kornat_helpful", "acc_norm,none", "Ko-Helpfulness") # 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)]) auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)]) auto_eval_column_dict.append(["flagged", ColumnContent, ColumnContent("Flagged", "bool", False, False)]) # Dummy column for the search bar (hidden by the custom CSS) auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)]) # We use make dataclass to dynamically fill the scores from Tasks AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True) @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) # Define the human baselines human_baseline_row = { AutoEvalColumn.model.name: "

Human performance

", } @dataclass class ModelDetails: name: str symbol: str = "" # emoji, only for the model type class ModelType(Enum): PT = ModelDetails(name="pretrained", symbol="🟢") CPT = ModelDetails(name="continuously pretrained", symbol="🟩") FT = ModelDetails(name="fine-tuned on domain-specific datasets", symbol="🔶") chat = ModelDetails(name="chat models (RLHF, DPO, IFT, ...)", symbol="💬") merges = ModelDetails(name="base merges and moerges", symbol="🤝") Unknown = ModelDetails(name="other", symbol="❓") def to_str(self, separator=" "): return f"{self.value.symbol}{separator}{self.value.name}" @staticmethod def from_str(m_type): if any([k for k in m_type if k in ["fine-tuned","🔶", "finetuned"]]): return ModelType.FT if "continuously pretrained" in m_type or "🟩" in m_type: return ModelType.CPT if "pretrained" in m_type or "🟢" in m_type: return ModelType.PT if any([k in m_type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]): return ModelType.chat if "merge" in m_type or "🤝" in m_type: return ModelType.merges 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("?") @staticmethod 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 # Column selection COLS = [c.name for c in fields(AutoEvalColumn)] 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 = { "Unknown": pd.Interval(-1, 0, closed="right"), "0~3B": pd.Interval(0, 3, closed="right"), "3~7B": pd.Interval(3, 7.3, closed="right"), "7~13B": pd.Interval(7.3, 13, closed="right"), "13~35B": pd.Interval(13, 35, closed="right"), "35~60B": pd.Interval(35, 60, closed="right"), "60B+": pd.Interval(60, 10000, closed="right"), }