from __future__ import annotations import json import os import re from functools import reduce from typing import Any import pandas as pd from datasets import load_dataset from huggingface_hub import hf_hub_download from huggingface_hub.repocard import metadata_load from tqdm.autonotebook import tqdm from envs import API, LEADERBOARD_CONFIG, MODEL_META, REPO_ID, RESULTS_REPO from utils.model_size import get_model_parameters_memory MODEL_CACHE = {} TASKS_CONFIG = LEADERBOARD_CONFIG["tasks"] BOARDS_CONFIG = LEADERBOARD_CONFIG["boards"] TASKS = list(TASKS_CONFIG.keys()) PRETTY_NAMES = { "InstructionRetrieval": "Retrieval w/Instructions", "PairClassification": "Pair Classification", "BitextMining": "Bitext Mining", } TASK_TO_METRIC = {k: [v["metric"]] for k, v in TASKS_CONFIG.items()} # Add legacy metric names TASK_TO_METRIC["STS"].append("cos_sim_spearman") TASK_TO_METRIC["STS"].append("cosine_spearman") TASK_TO_METRIC["Summarization"].append("cos_sim_spearman") TASK_TO_METRIC["Summarization"].append("cosine_spearman") TASK_TO_METRIC["PairClassification"].append("cos_sim_ap") TASK_TO_METRIC["PairClassification"].append("cosine_ap") EXTERNAL_MODELS = { k for k, v in MODEL_META["model_meta"].items() if v.get("is_external", False) } EXTERNAL_MODEL_TO_LINK = { k: v["link"] for k, v in MODEL_META["model_meta"].items() if v.get("link", False) } EXTERNAL_MODEL_TO_DIM = { k: v["dim"] for k, v in MODEL_META["model_meta"].items() if v.get("dim", False) } EXTERNAL_MODEL_TO_SEQLEN = { k: v["seq_len"] for k, v in MODEL_META["model_meta"].items() if v.get("seq_len", False) } EXTERNAL_MODEL_TO_SIZE = { k: v["size"] for k, v in MODEL_META["model_meta"].items() if v.get("size", False) } PROPRIETARY_MODELS = { k for k, v in MODEL_META["model_meta"].items() if v.get("is_proprietary", False) } TASK_DESCRIPTIONS = {k: v["task_description"] for k, v in TASKS_CONFIG.items()} TASK_DESCRIPTIONS["Overall"] = "Overall performance across MTEB tasks." SENTENCE_TRANSFORMERS_COMPATIBLE_MODELS = { k for k, v in MODEL_META["model_meta"].items() if v.get("is_sentence_transformers_compatible", False) } MODELS_TO_SKIP = MODEL_META["models_to_skip"] CROSS_ENCODERS = MODEL_META["cross_encoders"] BI_ENCODERS = [ k for k, _ in MODEL_META["model_meta"].items() if k not in CROSS_ENCODERS + ["bm25"] ] INSTRUCT_MODELS = { k for k, v in MODEL_META["model_meta"].items() if v.get("uses_instruct", False) } NOINSTRUCT_MODELS = { k for k, v in MODEL_META["model_meta"].items() if not v.get("uses_instruct", False) } TASK_TO_TASK_TYPE = {task_category: [] for task_category in TASKS} for board_config in BOARDS_CONFIG.values(): for task_category, task_list in board_config["tasks"].items(): TASK_TO_TASK_TYPE[task_category].extend(task_list) ## Don't cache this because we want to re-compute every time # model_infos_path = "model_infos.json" MODEL_INFOS = {} # if os.path.exists(model_infos_path): # with open(model_infos_path) as f: # MODEL_INFOS = json.load(f) def add_rank(df: pd.DataFrame) -> pd.DataFrame: cols_to_rank = [ col for col in df.columns if col not in [ "Model", "Model Size (Million Parameters)", "Memory Usage (GB, fp32)", "Embedding Dimensions", "Max Tokens", ] ] if len(cols_to_rank) == 1: df.sort_values(cols_to_rank[0], ascending=False, inplace=True) else: df.insert( len(df.columns) - len(cols_to_rank), "Average", df[cols_to_rank].mean(axis=1, skipna=False), ) df.sort_values("Average", ascending=False, inplace=True) df.insert(0, "Rank", list(range(1, len(df) + 1))) df = df.round(2) # Fill NaN after averaging df.fillna("", inplace=True) return df def make_clickable_model(model_name: str, link: None | str = None) -> str: if link is None: link = "https://huggingface.co/" + model_name # Remove user from model name return f'{model_name.split("/")[-1]}' def add_lang(examples): if not (examples["eval_language"]) or (examples["eval_language"] == "default"): examples["mteb_dataset_name_with_lang"] = examples["mteb_dataset_name"] else: examples["mteb_dataset_name_with_lang"] = ( examples["mteb_dataset_name"] + f' ({examples["eval_language"]})' ) return examples def norm(names: str) -> set: return set([name.split(" ")[0] for name in names]) def add_task(examples): # Could be added to the dataset loading script instead task_name = examples["mteb_dataset_name"] task_type = None for task_category, task_list in TASK_TO_TASK_TYPE.items(): if task_name in norm(task_list): task_type = task_category break if task_type is not None: examples["mteb_task"] = task_type else: print("WARNING: Task not found for dataset", examples["mteb_dataset_name"]) examples["mteb_task"] = "Unknown" return examples def filter_metric_external(x, task, metrics) -> bool: # This is a hack for the passkey and needle retrieval test, which reports ndcg_at_1 (i.e. accuracy), rather than the ndcg_at_10 that is commonly used for retrieval tasks. if x["mteb_dataset_name"] in ["LEMBNeedleRetrieval", "LEMBPasskeyRetrieval"]: return bool(x["mteb_task"] == task and x["metric"] == "ndcg_at_1") else: return bool(x["mteb_task"] == task and x["metric"] in metrics) def filter_metric_fetched(name: str, metric: str, expected_metrics) -> bool: # This is a hack for the passkey and needle retrieval test, which reports ndcg_at_1 (i.e. accuracy), rather than the ndcg_at_10 that is commonly used for retrieval tasks. return bool( metric == "ndcg_at_1" if name in ["LEMBNeedleRetrieval", "LEMBPasskeyRetrieval"] else metric in expected_metrics ) def get_dim_seq_size(model): filenames = [sib.rfilename for sib in model.siblings] dim, seq = "", "" for filename in filenames: if re.match("\d+_Pooling/config.json", filename): st_config_path = hf_hub_download(model.modelId, filename=filename) dim = json.load(open(st_config_path)).get("word_embedding_dimension", "") break for filename in filenames: if re.match("\d+_Dense/config.json", filename): st_config_path = hf_hub_download(model.modelId, filename=filename) dim = json.load(open(st_config_path)).get("out_features", dim) if "config.json" in filenames: config_path = hf_hub_download(model.modelId, filename="config.json") config = json.load(open(config_path)) if not dim: dim = config.get( "hidden_dim", config.get("hidden_size", config.get("d_model", "")) ) seq = config.get( "n_positions", config.get( "max_position_embeddings", config.get("n_ctx", config.get("seq_length", "")), ), ) if dim == "" or seq == "": raise Exception(f"Could not find dim or seq for model {model.modelId}") # Get model file size without downloading. Parameters in million parameters and memory in GB parameters, memory = get_model_parameters_memory(model) return dim, seq, parameters, memory def get_external_model_results(): if os.path.exists("EXTERNAL_MODEL_RESULTS.json"): with open("EXTERNAL_MODEL_RESULTS.json") as f: EXTERNAL_MODEL_RESULTS = json.load(f) # Update with models not contained models_to_run = [] for model in EXTERNAL_MODELS: if model not in EXTERNAL_MODEL_RESULTS: models_to_run.append(model) EXTERNAL_MODEL_RESULTS[model] = { k: {v[0]: []} for k, v in TASK_TO_METRIC.items() } ## only if we want to re-calculate all instead of using the cache... it's likely they haven't changed ## but if your model results have changed, delete it from the "EXTERNAL_MODEL_RESULTS.json" file else: EXTERNAL_MODEL_RESULTS = { model: {k: {v[0]: []} for k, v in TASK_TO_METRIC.items()} for model in EXTERNAL_MODELS } models_to_run = EXTERNAL_MODELS pbar = tqdm(models_to_run, desc="Fetching external model results") for model in pbar: pbar.set_description(f"Fetching external model results for {model!r}") ds = load_dataset( RESULTS_REPO, model, trust_remote_code=True, download_mode="force_redownload", verification_mode="no_checks", ) ds = ds.map(add_lang) ds = ds.map(add_task) base_dict = { "Model": make_clickable_model( model, link=EXTERNAL_MODEL_TO_LINK.get( model, f"https://huggingface.co/spaces/{REPO_ID}" ), ) } for task, metrics in TASK_TO_METRIC.items(): ds_dict = ds.filter(lambda x: filter_metric_external(x, task, metrics))[ "test" ].to_dict() ds_dict = { k: round(v, 2) for k, v in zip( ds_dict["mteb_dataset_name_with_lang"], ds_dict["score"] ) } # metrics[0] is the main name for this metric; other names in the list are legacy for backward-compat EXTERNAL_MODEL_RESULTS[model][task][metrics[0]].append( {**base_dict, **ds_dict} ) # Save & cache EXTERNAL_MODEL_RESULTS with open("EXTERNAL_MODEL_RESULTS.json", "w") as f: json.dump(EXTERNAL_MODEL_RESULTS, f, indent=4) return EXTERNAL_MODEL_RESULTS def download_or_use_cache(modelId: str): global MODEL_CACHE if modelId in MODEL_CACHE: return MODEL_CACHE[modelId] try: readme_path = hf_hub_download(modelId, filename="README.md", etag_timeout=30) except Exception: print(f"ERROR: Could not fetch metadata for {modelId}, trying again") readme_path = hf_hub_download(modelId, filename="README.md", etag_timeout=30) meta = metadata_load(readme_path) MODEL_CACHE[modelId] = meta return meta def get_mteb_data( tasks: list = ["Clustering"], langs: list = [], datasets: list = [], fillna: bool = True, add_emb_dim: bool = True, task_to_metric: dict = TASK_TO_METRIC, rank: bool = True, ) -> pd.DataFrame: global MODEL_INFOS with open("EXTERNAL_MODEL_RESULTS.json", "r") as f: external_model_results = json.load(f) api = API models = list(api.list_models(filter="mteb")) # Legacy names changes; Also fetch the old results & merge later if "MLSUMClusteringP2P (fr)" in datasets: datasets.append("MLSUMClusteringP2P") if "MLSUMClusteringS2S (fr)" in datasets: datasets.append("MLSUMClusteringS2S") if "PawsXPairClassification (fr)" in datasets: datasets.append("PawsX (fr)") # Initialize list to models that we cannot fetch metadata from df_list = [] for model in external_model_results: results_list = [] for task in tasks: # Not all models have InstructionRetrieval, other new tasks if task not in external_model_results[model]: continue results_list += external_model_results[model][task][task_to_metric[task][0]] if len(datasets) > 0: res = { k: v for d in results_list for k, v in d.items() if (k == "Model") or any([x in k for x in datasets]) } elif langs: # Would be cleaner to rely on an extra language column instead langs_format = [f"({lang})" for lang in langs] res = { k: v for d in results_list for k, v in d.items() if any([k.split(" ")[-1] in (k, x) for x in langs_format]) } else: res = {k: v for d in results_list for k, v in d.items()} # Model & at least one result if len(res) > 1: if add_emb_dim: res["Model Size (Million Parameters)"] = EXTERNAL_MODEL_TO_SIZE.get( model, "" ) res["Memory Usage (GB, fp32)"] = ( round(res["Model Size (Million Parameters)"] * 1e6 * 4 / 1024**3, 2) if res["Model Size (Million Parameters)"] != "" else "" ) res["Embedding Dimensions"] = EXTERNAL_MODEL_TO_DIM.get(model, "") res["Max Tokens"] = EXTERNAL_MODEL_TO_SEQLEN.get(model, "") df_list.append(res) pbar = tqdm(models, desc="Fetching model metadata") for model in pbar: if model.modelId in MODELS_TO_SKIP: continue if "gguf" in model.tags: continue pbar.set_description(f"Fetching {model.modelId!r} metadata") meta = download_or_use_cache(model.modelId) MODEL_INFOS[model.modelId] = {"metadata": meta} if "model-index" not in meta: continue # meta['model-index'][0]["results"] is list of elements like: # { # "task": {"type": "Classification"}, # "dataset": { # "type": "mteb/amazon_massive_intent", # "name": "MTEB MassiveIntentClassification (nb)", # "config": "nb", # "split": "test", # }, # "metrics": [ # {"type": "accuracy", "value": 39.81506388702084}, # {"type": "f1", "value": 38.809586587791664}, # ], # }, # Use "get" instead of dict indexing to skip incompat metadata instead of erroring out if len(datasets) > 0: task_results = [ sub_res for sub_res in meta["model-index"][0]["results"] if (sub_res.get("task", {}).get("type", "") in tasks) and any( [x in sub_res.get("dataset", {}).get("name", "") for x in datasets] ) ] elif langs: task_results = [ sub_res for sub_res in meta["model-index"][0]["results"] if (sub_res.get("task", {}).get("type", "") in tasks) and ( sub_res.get("dataset", {}).get("config", "default") in ("default", *langs) ) ] else: task_results = [ sub_res for sub_res in meta["model-index"][0]["results"] if (sub_res.get("task", {}).get("type", "") in tasks) ] try: out = [ { res["dataset"]["name"].replace("MTEB ", ""): [ round(score["value"], 2) for score in res["metrics"] if filter_metric_fetched( res["dataset"]["name"].replace("MTEB ", ""), score["type"], task_to_metric.get(res["task"]["type"]), ) ][0] } for res in task_results ] except Exception as e: print("ERROR", model.modelId, e) continue out = {k: v for d in out for k, v in d.items()} out["Model"] = make_clickable_model(model.modelId) # Model & at least one result if len(out) > 1: if add_emb_dim: # The except clause triggers on gated repos, we can use external metadata for those try: MODEL_INFOS[model.modelId]["dim_seq_size"] = list( get_dim_seq_size(model) ) except: name_without_org = model.modelId.split("/")[-1] # EXTERNAL_MODEL_TO_SIZE[name_without_org] refers to millions of parameters, so for memory usage # we multiply by 1e6 to get just the number of parameters, then by 4 to get the number of bytes # given fp32 precision (4 bytes per float), then divide by 1024**3 to get the number of GB MODEL_INFOS[model.modelId]["dim_seq_size"] = ( EXTERNAL_MODEL_TO_DIM.get(name_without_org, ""), EXTERNAL_MODEL_TO_SEQLEN.get(name_without_org, ""), EXTERNAL_MODEL_TO_SIZE.get(name_without_org, ""), round( EXTERNAL_MODEL_TO_SIZE[name_without_org] * 1e6 * 4 / 1024**3, 2, ) if name_without_org in EXTERNAL_MODEL_TO_SIZE else "", ) ( out["Embedding Dimensions"], out["Max Tokens"], out["Model Size (Million Parameters)"], out["Memory Usage (GB, fp32)"], ) = tuple(MODEL_INFOS[model.modelId]["dim_seq_size"]) df_list.append(out) model_siblings = model.siblings or [] if ( model.library_name == "sentence-transformers" or "sentence-transformers" in model.tags or "modules.json" in {file.rfilename for file in model_siblings} ): SENTENCE_TRANSFORMERS_COMPATIBLE_MODELS.add(out["Model"]) # # Save & cache MODEL_INFOS # with open("model_infos.json", "w") as f: # json.dump(MODEL_INFOS, f) df = pd.DataFrame(df_list) # If there are any models that are the same, merge them # E.g. if out["Model"] has the same value in two places, merge & take whichever one is not NaN else just take the first one df = df.groupby("Model", as_index=False).first() # Put 'Model' column first cols = sorted(list(df.columns)) base_columns = [ "Model", "Model Size (Million Parameters)", "Memory Usage (GB, fp32)", "Embedding Dimensions", "Max Tokens", ] if len(datasets) > 0: # Update legacy column names to be merged with newer ones # Update 'MLSUMClusteringP2P (fr)' with values from 'MLSUMClusteringP2P' if ("MLSUMClusteringP2P (fr)" in datasets) and ("MLSUMClusteringP2P" in cols): df["MLSUMClusteringP2P (fr)"] = df["MLSUMClusteringP2P (fr)"].fillna( df["MLSUMClusteringP2P"] ) datasets.remove("MLSUMClusteringP2P") if ("MLSUMClusteringS2S (fr)" in datasets) and ("MLSUMClusteringS2S" in cols): df["MLSUMClusteringS2S (fr)"] = df["MLSUMClusteringS2S (fr)"].fillna( df["MLSUMClusteringS2S"] ) datasets.remove("MLSUMClusteringS2S") if ("PawsXPairClassification (fr)" in datasets) and ("PawsX (fr)" in cols): # for the first bit no model has it, hence no column for it. We can remove this in a month or so if "PawsXPairClassification (fr)" not in cols: df["PawsXPairClassification (fr)"] = df["PawsX (fr)"] else: df["PawsXPairClassification (fr)"] = df[ "PawsXPairClassification (fr)" ].fillna(df["PawsX (fr)"]) # make all the columns the same datasets.remove("PawsX (fr)") cols.remove("PawsX (fr)") df.drop(columns=["PawsX (fr)"], inplace=True) # Filter invalid columns cols = [col for col in cols if col in base_columns + datasets] i = 0 for column in base_columns: if column in cols: cols.insert(i, cols.pop(cols.index(column))) i += 1 df = df[cols] if rank: df = add_rank(df) if fillna: df.fillna("", inplace=True) return df # Get dict with a task list for each task category # E.g. {"Classification": ["AmazonMassiveIntentClassification (en)", ...], "PairClassification": ["SprintDuplicateQuestions", ...]} def get_mteb_average(task_dict: dict) -> tuple[Any, dict]: all_tasks = reduce(lambda x, y: x + y, task_dict.values()) DATA_OVERALL = get_mteb_data( tasks=list(task_dict.keys()), datasets=all_tasks, fillna=False, add_emb_dim=True, rank=False, ) # Debugging: # DATA_OVERALL.to_csv("overall.csv") DATA_OVERALL.insert( 1, f"Average ({len(all_tasks)} datasets)", DATA_OVERALL[all_tasks].mean(axis=1, skipna=False), ) for i, (task_category, task_category_list) in enumerate(task_dict.items()): DATA_OVERALL.insert( i + 2, f"{task_category} Average ({len(task_category_list)} datasets)", DATA_OVERALL[task_category_list].mean(axis=1, skipna=False), ) DATA_OVERALL.sort_values( f"Average ({len(all_tasks)} datasets)", ascending=False, inplace=True ) # Start ranking from 1 DATA_OVERALL.insert(0, "Rank", list(range(1, len(DATA_OVERALL) + 1))) DATA_OVERALL = DATA_OVERALL.round(2) DATA_TASKS = {} for task_category, task_category_list in task_dict.items(): DATA_TASKS[task_category] = add_rank( DATA_OVERALL[ ["Model", "Model Size (Million Parameters)", "Memory Usage (GB, fp32)"] + task_category_list ] ) DATA_TASKS[task_category] = DATA_TASKS[task_category][ DATA_TASKS[task_category].iloc[:, 4:].ne("").any(axis=1) ] # Fill NaN after averaging DATA_OVERALL.fillna("", inplace=True) data_overall_rows = [ "Rank", "Model", "Model Size (Million Parameters)", "Memory Usage (GB, fp32)", "Embedding Dimensions", "Max Tokens", f"Average ({len(all_tasks)} datasets)", ] for task_category, task_category_list in task_dict.items(): data_overall_rows.append( f"{task_category} Average ({len(task_category_list)} datasets)" ) DATA_OVERALL = DATA_OVERALL[data_overall_rows] DATA_OVERALL = DATA_OVERALL[DATA_OVERALL.iloc[:, 5:].ne("").any(axis=1)] return DATA_OVERALL, DATA_TASKS def refresh_leaderboard() -> tuple[list, dict]: """ The main code to refresh and calculate results for the leaderboard. It does this by fetching the results from the external models and the models in the leaderboard, then calculating the average scores for each task category. """ # get external model results and cache them # NOTE: if your model results have changed, use this function to refresh them (see inside for details) get_external_model_results() boards_data = {} all_data_tasks = [] pbar_tasks = tqdm( BOARDS_CONFIG.items(), desc="Fetching leaderboard results for ???", total=len(BOARDS_CONFIG), leave=True, ) for board, board_config in pbar_tasks: boards_data[board] = {"data_overall": None, "data_tasks": {}} pbar_tasks.set_description(f"Fetching leaderboard results for {board!r}") pbar_tasks.refresh() if board_config["has_overall"]: data_overall, data_tasks = get_mteb_average(board_config["tasks"]) boards_data[board]["data_overall"] = data_overall boards_data[board]["data_tasks"] = data_tasks all_data_tasks.extend(data_tasks.values()) else: for task_category, task_category_list in board_config["tasks"].items(): data_task_category = get_mteb_data( tasks=[task_category], datasets=task_category_list ) data_task_category.drop( columns=["Embedding Dimensions", "Max Tokens"], inplace=True ) boards_data[board]["data_tasks"][task_category] = data_task_category all_data_tasks.append(data_task_category) return all_data_tasks, boards_data def write_out_results(item: dict, item_name: str) -> None: """ Due to their complex structure, let's recursively create subfolders until we reach the end of the item and then save the DFs as jsonl files Args: item: The item to save item_name: The name of the item """ main_folder = item_name if isinstance(item, list): for i, v in enumerate(item): write_out_results(v, os.path.join(main_folder, str(i))) elif isinstance(item, dict): for key, value in item.items(): if isinstance(value, dict): write_out_results(value, os.path.join(main_folder, key)) elif isinstance(value, list): for i, v in enumerate(value): write_out_results(v, os.path.join(main_folder, key + str(i))) else: write_out_results(value, os.path.join(main_folder, key)) elif isinstance(item, pd.DataFrame): print(f"Saving {main_folder} to {main_folder}/default.jsonl") os.makedirs(main_folder, exist_ok=True) if "index" not in item.columns: item.reset_index(inplace=True) item.to_json(f"{main_folder}/default.jsonl", orient="records", lines=True) elif isinstance(item, str): print(f"Saving {main_folder} to {main_folder}/default.txt") os.makedirs(main_folder, exist_ok=True) with open(f"{main_folder}/default.txt", "w") as f: f.write(item) elif item is None: # write out an empty file print(f"Saving {main_folder} to {main_folder}/default.txt") os.makedirs(main_folder, exist_ok=True) with open(f"{main_folder}/default.txt", "w") as f: f.write("") else: raise Exception(f"Unknown type {type(item)}") def load_results(data_path: str) -> list | dict | pd.DataFrame | str | None: """ Do the reverse of `write_out_results` to reconstruct the item Args: data_path: The path to the data to load Returns: The loaded data """ if os.path.isdir(data_path): # if the folder just has numbers from 0 to N, load as a list all_files_in_dir = list(os.listdir(data_path)) if set(all_files_in_dir) == set([str(i) for i in range(len(all_files_in_dir))]): ### the list case return [ load_results(os.path.join(data_path, str(i))) for i in range(len(os.listdir(data_path))) ] else: if len(all_files_in_dir) == 1: file_name = all_files_in_dir[0] if file_name == "default.jsonl": return load_results(os.path.join(data_path, file_name)) else: ### the dict case return {file_name: load_results(os.path.join(data_path, file_name))} else: return { file_name: load_results(os.path.join(data_path, file_name)) for file_name in all_files_in_dir } elif data_path.endswith(".jsonl"): df = pd.read_json(data_path, orient="records", lines=True) if "index" in df.columns: df = df.set_index("index") return df else: with open(data_path, "r") as f: data = f.read() if data == "": return None else: return data if __name__ == "__main__": print("Refreshing leaderboard statistics...") all_data_tasks, boards_data = refresh_leaderboard() print("Done calculating, saving...") # save them so that the leaderboard can use them. They're quite complex though # but we can't use pickle files because of git-lfs. write_out_results(all_data_tasks, "all_data_tasks") write_out_results(boards_data, "boards_data") # to load them use # all_data_tasks = load_results("all_data_tasks") # boards_data = load_results("boards_data") print("Done saving results!")