Duplicate from HuggingFaceH4/open_llm_leaderboard

Co-authored-by: Edward Beeching <edbeeching@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,34 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,3 @@
+evals/
+venv/
+__pycache__/

README.md

@@ -0,0 +1,14 @@
+---
+title: Open LLM Leaderboard
+emoji: 👀
+colorFrom: green
+colorTo: indigo
+sdk: gradio
+sdk_version: 3.27.0
+app_file: app.py
+pinned: false
+license: apache-2.0
+duplicated_from: HuggingFaceH4/open_llm_leaderboard
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,256 @@
+import os
+import shutil
+import numpy as np
+import gradio as gr
+from huggingface_hub import Repository, HfApi
+from transformers import AutoConfig
+import json
+from apscheduler.schedulers.background import BackgroundScheduler
+import pandas as pd
+import datetime
+from utils import get_eval_results_dicts, make_clickable_model
+
+# clone / pull the lmeh eval data
+H4_TOKEN = os.environ.get("H4_TOKEN", None)
+LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"
+IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", None))
+
+repo=None
+if H4_TOKEN:
+    print("pulling repo")
+    # try:
+    #     shutil.rmtree("./evals/")
+    # except:
+    #     pass
+
+    repo = Repository(
+        local_dir="./evals/", clone_from=LMEH_REPO, use_auth_token=H4_TOKEN, repo_type="dataset"
+    )
+    repo.git_pull()
+
+
+# parse the results
+BENCHMARKS = ["arc_challenge", "hellaswag", "hendrycks", "truthfulqa_mc"]
+
+METRICS = ["acc_norm", "acc_norm", "acc_norm", "mc2"]
+
+
+def load_results(model, benchmark, metric):
+    file_path = os.path.join("evals", model, f"{model}-eval_{benchmark}.json")
+    if not os.path.exists(file_path):
+        return 0.0, None
+
+    with open(file_path) as fp:
+        data = json.load(fp)
+    accs = np.array([v[metric] for k, v in data["results"].items()])
+    mean_acc = np.mean(accs)  
+    return mean_acc, data["config"]["model_args"]
+
+
+COLS = ["Model", "Revision", "Average ⬆️", "ARC (25-shot) ⬆️", "HellaSwag (10-shot) ⬆️", "MMLU (5-shot) ⬆️", "TruthQA (0-shot) ⬆️"]
+TYPES = ["markdown","str", "number", "number", "number", "number", "number", ]
+
+if not IS_PUBLIC:
+    COLS.insert(2, "8bit")
+    TYPES.insert(2, "bool")
+
+EVAL_COLS = ["model", "revision", "private", "8bit_eval", "is_delta_weight", "status"]
+EVAL_TYPES = ["markdown","str", "bool", "bool", "bool", "str"]
+def get_leaderboard():
+    if repo: 
+        print("pulling changes")
+        repo.git_pull()
+        
+    all_data = get_eval_results_dicts(IS_PUBLIC)
+    
+    if not IS_PUBLIC:
+        gpt4_values = {
+            "Model":f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: blue; text-decoration: underline;text-decoration-style: dotted;">gpt4</a>', 
+            "Revision":"tech report", 
+            "8bit":None,
+            "Average ⬆️":84.3,
+            "ARC (25-shot) ⬆️":96.3,
+            "HellaSwag (10-shot) ⬆️":95.3,
+            "MMLU (5-shot) ⬆️":86.4,
+            "TruthQA (0-shot) ⬆️":59.0,
+        }
+        all_data.append(gpt4_values)
+        gpt35_values = {
+            "Model":f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: blue; text-decoration: underline;text-decoration-style: dotted;">gpt3.5</a>', 
+            "Revision":"tech report", 
+            "8bit":None,
+            "Average ⬆️":71.9,
+            "ARC (25-shot) ⬆️":85.2,
+            "HellaSwag (10-shot) ⬆️":85.5,
+            "MMLU (5-shot) ⬆️":70.0,
+            "TruthQA (0-shot) ⬆️":47.0,
+        }
+        all_data.append(gpt35_values)
+    
+    dataframe = pd.DataFrame.from_records(all_data)
+    dataframe = dataframe.sort_values(by=['Average ⬆️'], ascending=False)
+    print(dataframe)
+    dataframe = dataframe[COLS]
+    return dataframe
+
+def get_eval_table():
+    if repo: 
+        print("pulling changes for eval")
+        repo.git_pull()
+    entries = [entry for entry in os.listdir("evals/eval_requests") if not entry.startswith('.')] 
+    all_evals = []
+    
+    for entry in entries:
+        print(entry)
+        if ".json"in entry:
+            file_path = os.path.join("evals/eval_requests", entry)
+            with open(file_path) as fp:
+                data = json.load(fp)
+                
+            data["# params"] = "unknown"
+            data["model"] = make_clickable_model(data["model"])
+            data["revision"] = data.get("revision", "main")
+            
+
+            all_evals.append(data)
+        else:
+            # this is a folder
+            sub_entries = [e for e in os.listdir(f"evals/eval_requests/{entry}") if not e.startswith('.')] 
+            for sub_entry in sub_entries:
+                file_path = os.path.join("evals/eval_requests", entry, sub_entry)
+                with open(file_path) as fp:
+                    data = json.load(fp)
+                    
+                #data["# params"] = get_n_params(data["model"])
+                data["model"] = make_clickable_model(data["model"])
+                all_evals.append(data)
+
+    
+    dataframe = pd.DataFrame.from_records(all_evals)
+    return dataframe[EVAL_COLS]
+
+
+leaderboard = get_leaderboard()
+eval_queue = get_eval_table()
+
+def is_model_on_hub(model_name, revision) -> bool:
+    try:
+        config = AutoConfig.from_pretrained(model_name, revision=revision)
+        return True
+        
+    except Exception as e:
+        print("Could not get the model config from the hub")
+        print(e)
+        return False
+        
+
+
+def add_new_eval(model:str, base_model : str, revision:str, is_8_bit_eval: bool, private:bool, is_delta_weight:bool):
+    # check the model actually exists before adding the eval
+    if revision == "":
+        revision = "main"
+    if is_delta_weight and not is_model_on_hub(base_model, revision):
+        print(base_model, "base model not found on hub")
+        return
+    
+    if not is_model_on_hub(model, revision):
+        print(model, "not found on hub")
+        return
+    print("adding new eval")
+    
+    eval_entry = {
+        "model" : model,
+        "base_model" : base_model,
+        "revision" : revision,
+        "private" : private,
+        "8bit_eval" : is_8_bit_eval,
+        "is_delta_weight" : is_delta_weight,
+        "status" : "PENDING"
+    }    
+    
+    user_name = ""
+    model_path = model
+    if "/" in model:
+        user_name = model.split("/")[0]
+        model_path = model.split("/")[1]
+    
+    OUT_DIR=f"eval_requests/{user_name}"
+    os.makedirs(OUT_DIR, exist_ok=True)
+    out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{is_8_bit_eval}_{is_delta_weight}.json"
+    
+    with open(out_path, "w") as f:
+        f.write(json.dumps(eval_entry))
+    LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"
+    
+    api = HfApi()
+    api.upload_file(
+        path_or_fileobj=out_path,
+        path_in_repo=out_path,
+        repo_id=LMEH_REPO,
+        token=H4_TOKEN,
+        repo_type="dataset",
+    )
+
+    
+def refresh():
+    return get_leaderboard(), get_eval_table()
+    
+
+
+block = gr.Blocks()
+with block: 
+    with gr.Row():
+        gr.Markdown(f"""
+# 🤗 Open LLM Leaderboard
+<font size="4">With the plethora of large language models (LLMs) and chatbots being released week upon week, often with grandiose claims of their performance, it can be hard to filter out the genuine progress that is being made by the open-source community and which model is the current state of the art. The 🤗 Open LLM Leaderboard aims to track, rank and evaluate LLMs and chatbots as they are released. We evaluate models on 4 key benchmarks from the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks. A key advantage of this leaderboard is that anyone from the community can submit a model for automated evaluation on the 🤗 GPU cluster, as long as it is a 🤗 Transformers model with weights on the Hub. We also support evaluation of models with delta-weights for non-commercial licensed models, such as LLaMa.
+
+Evaluation is performed against 4 popular benchmarks:
+- <a href="https://arxiv.org/abs/1803.05457" target="_blank">  AI2 Reasoning Challenge </a> (25-shot) - a set of grade-school science questions.
+- <a href="https://arxiv.org/abs/1905.07830" target="_blank">  HellaSwag </a> (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models.
+- <a href="https://arxiv.org/abs/2009.03300" target="_blank">  MMLU </a>  (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
+- <a href="https://arxiv.org/abs/2109.07958" target="_blank">  Truthful QA MC </a> (0-shot) - a benchmark to measure whether a language model is truthful in generating answers to questions. 
+
+We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings. </font>
+        """)
+    
+    with gr.Row():
+        leaderboard_table = gr.components.Dataframe(value=leaderboard, headers=COLS,
+                                                    datatype=TYPES, max_rows=5)
+
+    
+    
+    with gr.Row():
+        gr.Markdown(f"""
+    # Evaluation Queue for the 🤗 Open LLM Leaderboard, these models will be automatically evaluated on the 🤗 cluster
+    
+    """)
+    with gr.Accordion("Evaluation Queue", open=False):
+        with gr.Row():
+            eval_table = gr.components.Dataframe(value=eval_queue, headers=EVAL_COLS,
+                                                        datatype=EVAL_TYPES, max_rows=5)    
+        
+    with gr.Row():
+        refresh_button = gr.Button("Refresh")
+        refresh_button.click(refresh, inputs=[], outputs=[leaderboard_table, eval_table]) 
+        
+    with gr.Accordion("Submit a new model for evaluation"):
+        # with gr.Row():
+        #     gr.Markdown(f"""# Submit a new model for evaluation""")
+        with gr.Row():
+            with gr.Column():
+                model_name_textbox = gr.Textbox(label="Model name")
+                revision_name_textbox = gr.Textbox(label="revision", placeholder="main")
+                
+            with gr.Column():
+                is_8bit_toggle = gr.Checkbox(False, label="8 bit eval", visible=not IS_PUBLIC)
+                private = gr.Checkbox(False, label="Private", visible=not IS_PUBLIC)
+                is_delta_weight = gr.Checkbox(False, label="Delta weights")
+                base_model_name_textbox = gr.Textbox(label="base model (for delta)")
+            
+        with gr.Row():
+            submit_button = gr.Button("Submit Eval")
+            submit_button.click(add_new_eval, [model_name_textbox, base_model_name_textbox, revision_name_textbox, is_8bit_toggle, private, is_delta_weight])
+        
+
+    block.load(refresh, inputs=[], outputs=[leaderboard_table, eval_table])
+block.launch()

requirements.txt

@@ -0,0 +1,67 @@
+aiofiles==23.1.0
+aiohttp==3.8.4
+aiosignal==1.3.1
+altair==4.2.2
+anyio==3.6.2
+APScheduler==3.10.1
+async-timeout==4.0.2
+attrs==23.1.0
+certifi==2022.12.7
+charset-normalizer==3.1.0
+click==8.1.3
+contourpy==1.0.7
+cycler==0.11.0
+entrypoints==0.4
+fastapi==0.95.1
+ffmpy==0.3.0
+filelock==3.11.0
+fonttools==4.39.3
+frozenlist==1.3.3
+fsspec==2023.4.0
+gradio==3.27.0
+gradio_client==0.1.3
+h11==0.14.0
+httpcore==0.17.0
+httpx==0.24.0
+huggingface-hub==0.13.4
+idna==3.4
+Jinja2==3.1.2
+jsonschema==4.17.3
+kiwisolver==1.4.4
+linkify-it-py==2.0.0
+markdown-it-py==2.2.0
+MarkupSafe==2.1.2
+matplotlib==3.7.1
+mdit-py-plugins==0.3.3
+mdurl==0.1.2
+multidict==6.0.4
+numpy==1.24.2
+orjson==3.8.10
+packaging==23.1
+pandas==2.0.0
+Pillow==9.5.0
+pydantic==1.10.7
+pydub==0.25.1
+pyparsing==3.0.9
+pyrsistent==0.19.3
+python-dateutil==2.8.2
+python-multipart==0.0.6
+pytz==2023.3
+pytz-deprecation-shim==0.1.0.post0
+PyYAML==6.0
+requests==2.28.2
+semantic-version==2.10.0
+six==1.16.0
+sniffio==1.3.0
+starlette==0.26.1
+toolz==0.12.0
+tqdm==4.65.0
+transformers==4.28.1
+typing_extensions==4.5.0
+tzdata==2023.3
+tzlocal==4.3
+uc-micro-py==1.0.1
+urllib3==1.26.15
+uvicorn==0.21.1
+websockets==11.0.1
+yarl==1.8.2

utils.py

@@ -0,0 +1,136 @@
+import os
+import shutil
+import numpy as np
+import gradio as gr
+from huggingface_hub import Repository, HfApi
+from transformers import AutoConfig, AutoModel
+import json
+from apscheduler.schedulers.background import BackgroundScheduler
+import pandas as pd
+import datetime
+import glob
+from dataclasses import dataclass
+from typing import List, Tuple, Dict
+# clone / pull the lmeh eval data
+H4_TOKEN = os.environ.get("H4_TOKEN", None)
+LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"
+
+METRICS = ["acc_norm", "acc_norm", "acc_norm", "mc2"]
+BENCHMARKS = ["arc_challenge", "hellaswag", "hendrycks", "truthfulqa_mc"]
+BENCH_TO_NAME = {
+    "arc_challenge":"ARC (25-shot) ⬆️",
+     "hellaswag":"HellaSwag (10-shot) ⬆️",
+     "hendrycks":"MMLU (5-shot) ⬆️",
+     "truthfulqa_mc":"TruthQA (0-shot) ⬆️",
+}
+def make_clickable_model(model_name):        
+    LLAMAS = ["huggingface/llama-7b", "huggingface/llama-13b", "huggingface/llama-30b", "huggingface/llama-65b"]
+    if model_name in LLAMAS:
+        model = model_name.split("/")[1]
+        return f'<a target="_blank" href="https://ai.facebook.com/blog/large-language-model-llama-meta-ai/" style="color: blue; text-decoration: underline;text-decoration-style: dotted;">{model}</a>'
+    
+    if model_name == "HuggingFaceH4/stable-vicuna-13b-2904":
+        link = "https://huggingface.co/" + "CarperAI/stable-vicuna-13b-delta"
+        return f'<a target="_blank" href="{link}" style="color: blue; text-decoration: underline;text-decoration-style: dotted;">stable-vicuna-13b</a>'
+    
+    if model_name == "HuggingFaceH4/llama-7b-ift-alpaca":
+        link = "https://crfm.stanford.edu/2023/03/13/alpaca.html"
+        return f'<a target="_blank" href="{link}" style="color: blue; text-decoration: underline;text-decoration-style: dotted;">alpaca-13b</a>'
+
+    # remove user from model name
+    #model_name_show = ' '.join(model_name.split('/')[1:])
+
+    link = "https://huggingface.co/" + model_name
+    return f'<a target="_blank" href="{link}" style="color: blue; text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
+
+@dataclass
+class EvalResult:
+    eval_name : str
+    org : str
+    model : str
+    revision : str
+    is_8bit : bool
+    results : dict
+    
+    def to_dict(self):
+        
+        if self.org is not None:
+            base_model =f"{self.org}/{self.model}"
+        else:
+            base_model =f"{self.model}"
+        data_dict = {}
+        
+        data_dict["eval_name"] = self.eval_name
+        data_dict["8bit"] = self.is_8bit
+        data_dict["Model"] = make_clickable_model(base_model)
+        data_dict["Revision"] = self.revision
+        data_dict["Average ⬆️"] = round(sum([v for k,v in self.results.items()])/4.0,1) 
+        #data_dict["# params"] = get_n_params(base_model)
+        
+        for benchmark in BENCHMARKS:
+            if not benchmark in self.results.keys():
+                self.results[benchmark] = None
+                
+        for k,v in BENCH_TO_NAME.items():
+            data_dict[v] = self.results[k]
+        
+        return data_dict
+        
+        
+   
+   
+def parse_eval_result(json_filepath: str) -> Tuple[str, dict]:
+    with open(json_filepath) as fp:
+        data = json.load(fp)
+
+    path_split = json_filepath.split("/")
+    org = None
+    model = path_split[-4]
+    is_8bit = path_split[-2] == "8bit"
+    revision = path_split[-3]
+    if len(path_split)== 7:
+        # handles gpt2 type models that don't have an org
+        result_key = f"{path_split[-4]}_{path_split[-3]}_{path_split[-2]}"
+    else:
+        result_key = f"{path_split[-5]}_{path_split[-4]}_{path_split[-3]}_{path_split[-2]}"
+        org = path_split[-5]
+        
+    eval_result = None
+    for benchmark, metric  in zip(BENCHMARKS, METRICS):
+        if benchmark in json_filepath:
+            accs = np.array([v[metric] for k, v in data["results"].items()])
+            mean_acc = round(np.mean(accs)*100.0,1)
+            eval_result = EvalResult(result_key, org, model, revision, is_8bit, {benchmark:mean_acc})
+        
+    return result_key, eval_result
+        
+    
+    
+   
+def get_eval_results(is_public) -> List[EvalResult]:
+    json_filepaths = glob.glob("evals/eval_results/public/**/16bit/*.json", recursive=True)
+    if not is_public:
+        json_filepaths += glob.glob("evals/eval_results/private/**/*.json", recursive=True)
+        json_filepaths += glob.glob("evals/eval_results/private/**/*.json", recursive=True)
+        json_filepaths += glob.glob("evals/eval_results/public/**/8bit/*.json", recursive=True) # include the 8bit evals of public models
+    eval_results = {}
+    
+    for json_filepath in json_filepaths:
+        result_key, eval_result = parse_eval_result(json_filepath)
+        if result_key in eval_results.keys():
+            eval_results[result_key].results.update(eval_result.results)
+        else:
+            eval_results[result_key] = eval_result
+        
+        
+    eval_results = [v for k,v in eval_results.items()]
+    
+    return eval_results
+    
+def get_eval_results_dicts(is_public=True) -> List[Dict]:
+    eval_results = get_eval_results(is_public)
+    
+    return [e.to_dict() for e in eval_results]
+
+eval_results_dict = get_eval_results_dicts()
+print(eval_results_dict)