from dataclasses import dataclass, make_dataclass, field from enum import Enum from typing import TypeVar import pandas as pd _E = TypeVar("_E", bound=Enum) def _enum_from_str(enum_cls: type[_E], value: str, default: _E) -> _E: """Generic enum lookup by value name. Returns *default* on miss.""" for member in enum_cls: if member.value.name == value: return member return default 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): arc = Task("arc:challenge", "acc,none", "ARC-c") arc_easy = Task("arc:easy", "acc,none", "ARC-e") boolq = Task("boolq", "acc,none", "Boolq") hellaswag = Task("hellaswag", "acc,none", "HellaSwag") lambada_openai = Task("lambada:openai", "acc,none", "Lambada") mmlu = Task("mmlu", "acc,none", "MMLU") openbookqa = Task("openbookqa", "acc,none", "Openbookqa") piqa = Task("piqa", "acc,none", "Piqa") # truthfulqa:mc1 / truthfulqa:mc2 -- ? truthfulqa_mc = Task("truthfulqa:mc1", "acc,none", "Truthfulqa") # arc:challenge ? # arc_challenge = Task("arc:challenge", "acc_norm,none", "Arc challenge") # truthfulqa = Task("truthfulqa:mc", "mc2", "TruthfulQA") winogrande = Task("winogrande", "acc,none", "Winogrande") # gsm8k = Task("gsm8k", "acc", "GSM8K") # 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_list = [] # Init auto_eval_column_list.append([ "model_type_symbol", ColumnContent, field(default_factory=lambda: ColumnContent("T", "str", False, hidden=True)) ]) auto_eval_column_list.append([ "model", ColumnContent, field(default_factory=lambda: ColumnContent("Model", "markdown", True, never_hidden=True)) ]) # Scores auto_eval_column_list.append([ "average", ColumnContent, field(default_factory=lambda: ColumnContent("Average", "number", True)) ]) for task in Tasks: auto_eval_column_list.append([ task.name, ColumnContent, field(default_factory=lambda t=task: ColumnContent(t.value.col_name, "number", True)) ]) auto_eval_column_list.append([ "params", ColumnContent, field(default_factory=lambda: ColumnContent("#Params (B)", "number", True)) ]) auto_eval_column_list.append([ "model_size", ColumnContent, field(default_factory=lambda: ColumnContent("#Size (G)", "number", True)) ]) # Dummy column for the search bar auto_eval_column_list.append([ "dummy", ColumnContent, field(default_factory=lambda: ColumnContent("model_name_for_query", "str", False, dummy=True)) ]) auto_eval_column_list.append(["model_type", ColumnContent, field(default_factory=lambda: ColumnContent("Type", "str", False, hidden=True))]) auto_eval_column_list.append(["architecture", ColumnContent, field(default_factory=lambda: ColumnContent("Architecture", "str", False))]) auto_eval_column_list.append(["weight_type", ColumnContent, field(default_factory=lambda: ColumnContent("Weight type", "str", False, True))]) auto_eval_column_list.append(["quant_type", ColumnContent, field(default_factory=lambda: ColumnContent("Quant type", "str", False))]) auto_eval_column_list.append(["precision", ColumnContent, field(default_factory=lambda: ColumnContent("Precision", "str", False))]) auto_eval_column_list.append(["weight_dtype", ColumnContent, field(default_factory=lambda: ColumnContent("Weight dtype", "str", False))]) auto_eval_column_list.append(["compute_dtype", ColumnContent, field(default_factory=lambda: ColumnContent("Compute dtype", "str", False))]) auto_eval_column_list.append(["merged", ColumnContent, field(default_factory=lambda: ColumnContent("Merged", "bool", False, hidden=True))]) auto_eval_column_list.append(["license", ColumnContent, field(default_factory=lambda: ColumnContent("Hub License", "str", False))]) auto_eval_column_list.append(["likes", ColumnContent, field(default_factory=lambda: ColumnContent("Hub ❤️", "number", False))]) auto_eval_column_list.append(["still_on_hub", ColumnContent, field(default_factory=lambda: ColumnContent("Available on the hub", "bool", False, hidden=True))]) auto_eval_column_list.append(["revision", ColumnContent, field(default_factory=lambda: ColumnContent("Model sha", "str", False, False))]) auto_eval_column_list.append(["flagged", ColumnContent, field(default_factory=lambda: ColumnContent("Flagged", "bool", False, hidden=True))]) auto_eval_column_list.append(["moe", ColumnContent, field(default_factory=lambda: ColumnContent("MoE", "bool", False, hidden=True))]) auto_eval_column_list.append(["double_quant", ColumnContent, field(default_factory=lambda: ColumnContent("Double Quant", "bool", False))]) auto_eval_column_list.append(["group_size", ColumnContent, field(default_factory=lambda: ColumnContent("Group Size", "bool", False))]) # We use make dataclass to dynamically fill the scores from Tasks # Fixed order: [model_type_symbol, model] + [model_size, params] + sorted rest _PINNED_AFTER_MODEL = {"model_size", "params"} _pinned = [x for x in auto_eval_column_list[2:] if x[0] in _PINNED_AFTER_MODEL] _rest = [x for x in auto_eval_column_list[2:] if x[0] not in _PINNED_AFTER_MODEL] sorted_columns = sorted(_rest, key=lambda x: x[0]) sorted_auto_eval_column_list = auto_eval_column_list[:2] + _pinned + sorted_columns AutoEvalColumn = make_dataclass("AutoEvalColumn", sorted_auto_eval_column_list, frozen=True) auto_eval_cols = AutoEvalColumn() @dataclass(frozen=True) class EvalQueueColumn: # Queue column for auto_eval 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", False) status = ColumnContent("status", "str", True) eta = ColumnContent("eta", "str", True) submitted_by = ColumnContent("submitted_by", "str", True) submitted_time = ColumnContent("submitted_time", "str", True) eval_queue_cols = EvalQueueColumn() @dataclass(frozen=True) class QuantQueueColumn: # Queue column for auto_quant model = ColumnContent("model", "markdown", True) revision = ColumnContent("revision", "str", True) private = ColumnContent("private", "bool", True) quant_scheme = ColumnContent("quant_scheme", "str", True) input_dtype = ColumnContent("input_dtype", "str", True) status = ColumnContent("status", "str", True) eta = ColumnContent("eta", "str", True) submitted_by = ColumnContent("submitted_by", "str", True) submitted_time = ColumnContent("submitted_time", "str", True) @dataclass class ModelDetails: name: str symbol: str = "" 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="", 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 "continously pretrained" in type or "🟩" in type: return ModelType.CPT if "pretrained" in type or "🟢" in type or "quantization" 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 return ModelType.Unknown class WeightType(Enum): Adapter = ModelDetails("Adapter") Original = ModelDetails("Original") Delta = ModelDetails("Delta") class QuantType(Enum): gptq = ModelDetails(name="GPTQ", symbol="🟢") aqlm = ModelDetails(name="AQLM", symbol="⭐") awq = ModelDetails(name="AWQ", symbol="🟩") llama_cpp = ModelDetails(name="llama.cpp", symbol="🔷") bnb = ModelDetails(name="bitsandbytes", symbol="🔵") autoround = ModelDetails(name="AutoRound", symbol="🍒") Unknown = ModelDetails(name="?", symbol="?") QuantType_None = ModelDetails(name="None", symbol="✖") def to_str(self, separator=" "): return f"{self.value.symbol}{separator}{self.value.name}" @staticmethod def from_str(quant_dtype): return _enum_from_str(QuantType, quant_dtype, QuantType.Unknown) class WeightDtype(Enum): all = ModelDetails("All") int2 = ModelDetails("int2") int3 = ModelDetails("int3") int4 = ModelDetails("int4") int8 = ModelDetails("int8") nf4 = ModelDetails("nf4") fp4 = ModelDetails("fp4") mxfp4 = ModelDetails("mxfp4") nvfp4 = ModelDetails("nvfp4") f16 = ModelDetails("float16") bf16 = ModelDetails("bfloat16") f32 = ModelDetails("float32") Unknown = ModelDetails("?") @staticmethod def from_str(weight_dtype): return _enum_from_str(WeightDtype, weight_dtype, WeightDtype.Unknown) class ComputeDtype(Enum): all = ModelDetails("All") fp16 = ModelDetails("float16") bf16 = ModelDetails("bfloat16") int8 = ModelDetails("int8") fp32 = ModelDetails("float32") Unknown = ModelDetails("?") @staticmethod def from_str(compute_dtype): return _enum_from_str(ComputeDtype, compute_dtype, ComputeDtype.Unknown) class GroupDtype(Enum): group_1 = ModelDetails("-1") group_1024 = ModelDetails("1024") group_256 = ModelDetails("256") group_128 = ModelDetails("128") group_64 = ModelDetails("64") group_32 = ModelDetails("32") group_all = ModelDetails("All") @staticmethod def from_str(group_dtype): return _enum_from_str(GroupDtype, group_dtype, GroupDtype.group_all) class Precision(Enum): # float16 = ModelDetails("float16") # bfloat16 = ModelDetails("bfloat16") qt_2bit = ModelDetails("2bit") qt_3bit = ModelDetails("3bit") qt_4bit = ModelDetails("4bit") qt_8bit = ModelDetails("8bit") qt_16bit = ModelDetails("16bit") qt_32bit = ModelDetails("32bit") Unknown = ModelDetails("?") @staticmethod def from_str(precision): return _enum_from_str(Precision, precision, Precision.Unknown) # Column selection COLS = [c.name for c in fields(auto_eval_cols)] TYPES = [c.type for c in fields(auto_eval_cols)] EVAL_COLS = [c.name for c in fields(EvalQueueColumn)] EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)] QUANT_COLS = [c.name for c in fields(QuantQueueColumn)] QUANT_TYPES = [c.type for c in fields(QuantQueueColumn)] # Display-only column lists (submitted_by, private, eta, revision hidden in UI) _QUEUE_UI_HIDDEN = {"submitted_by", "private", "eta", "revision"} QUANT_DISPLAY_COLS = [c for c in QUANT_COLS if c not in _QUEUE_UI_HIDDEN] QUANT_DISPLAY_TYPES = [t for c, t in zip(QUANT_COLS, QUANT_TYPES) if c not in _QUEUE_UI_HIDDEN] EVAL_DISPLAY_COLS = [c for c in EVAL_COLS if c not in _QUEUE_UI_HIDDEN] EVAL_DISPLAY_TYPES = [t for c, t in zip(EVAL_COLS, EVAL_TYPES) if c not in _QUEUE_UI_HIDDEN] # Human-readable column headers shown in queue tables. Underlying DataFrame # columns keep the original snake_case names; only the rendered header changes. _QUEUE_HEADER_LABELS = { "input_dtype": "weight_dtype", "submitted_time": "submitted time", } def _queue_headers(cols: list[str]) -> list[str]: return [_QUEUE_HEADER_LABELS.get(c, c) for c in cols] QUANT_DISPLAY_HEADERS = _queue_headers(QUANT_DISPLAY_COLS) EVAL_DISPLAY_HEADERS = _queue_headers(EVAL_DISPLAY_COLS) BENCHMARK_COLS = [t.value.col_name for t in Tasks] NUMERIC_INTERVALS = { "?": pd.Interval(-1, 0, closed="right"), "~1.5": pd.Interval(0, 2, closed="right"), "~3": pd.Interval(2, 4, closed="right"), "~7": pd.Interval(4, 9, closed="right"), "~13": pd.Interval(9, 20, closed="right"), "~35": pd.Interval(20, 45, closed="right"), "~60": pd.Interval(45, 70, closed="right"), "70+": pd.Interval(70, 10000, closed="right"), } NUMERIC_MODELSIZE = { "?": pd.Interval(-1, 0, closed="right"), "~4": pd.Interval(0, 4, closed="right"), "~8": pd.Interval(4, 8, closed="right"), "~16": pd.Interval(8, 16, closed="right"), "~36": pd.Interval(16, 36, closed="right"), "~48": pd.Interval(36, 48, closed="right"), "~64": pd.Interval(48, 64, closed="right"), ">72": pd.Interval(64, 200, closed="right"), }