added public option

app.py

@@ -8,11 +8,12 @@ import json
 from apscheduler.schedulers.background import BackgroundScheduler
 import pandas as pd
 import datetime
-from utils import get_eval_results_dicts, make_clickable_model, get_n_params
+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(True) # add secret here
 
 repo=None
 if H4_TOKEN:
@@ -46,8 +47,12 @@ def load_results(model, benchmark, metric):
     return mean_acc, data["config"]["model_args"]
 
 
-COLS = ["base_model", "revision", "8bit", "total ⬆️", "ARC (25-shot) ⬆️", "HellaSwag (10-shot) ⬆️", "MMLU (5-shot) ⬆️", "TruthQA (0-shot) ⬆️"]
-TYPES = ["markdown","str", "bool", "number", "number", "number", "number", "number", ]
+COLS = ["base_model", "revision", "total ⬆️", "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"]
@@ -56,7 +61,31 @@ def get_leaderboard():
         print("pulling changes")
         repo.git_pull()
         
-    all_data = get_eval_results_dicts()
+    all_data = get_eval_results_dicts(IS_PUBLIC)
+    
+    gpt4_values = {
+        "base_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,
+        "total ⬆️":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 = {
+        "base_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,
+        "total ⬆️":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=['total ⬆️'], ascending=False)
     print(dataframe)
@@ -77,7 +106,7 @@ def get_eval_table():
             with open(file_path) as fp:
                 data = json.load(fp)
                 
-            data["# params"] = get_n_params(data["model"])
+            data["# params"] = "unknown"
             data["model"] = make_clickable_model(data["model"])
             data["revision"] = data.get("revision", "main")
             
@@ -171,8 +200,16 @@ block = gr.Blocks()
 with block: 
     with gr.Row():
         gr.Markdown(f"""
-        # 🤗 H4 Model Evaluation leaderboard using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank"> LMEH benchmark suite </a>. 
-        Evaluation is performed against 4 popular benchmarks AI2 Reasoning Challenge, HellaSwag, MMLU, and TruthFul QC MC. To run your own benchmarks, refer to the README in the H4 repo.
+# 🤗 Open Chatbot Leaderboard
+<font size="4">With the plethora of chatbot LLMs 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 chatbot is the current state of the art. The 🤗 Open Chatbot Leaderboard aims to track, rank and evaluate chatbot models as they are released. We evaluate models of 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 🤗 research cluster. As long as it is Transformers model with weights on the 🤗 hub. We also support delta-weights for non-commercial licensed models, such as llama.
+<p>
+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. 
+<p>
+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():
@@ -186,10 +223,10 @@ with block:
     # Evaluation Queue for the LMEH benchmarks, these models will be automatically evaluated on the 🤗 cluster
     
     """)
-    
-    with gr.Row():
-        eval_table = gr.components.Dataframe(value=eval_queue, headers=EVAL_COLS,
-                                                    datatype=EVAL_TYPES, max_rows=5)    
+    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")
@@ -202,11 +239,12 @@ with block:
             with gr.Column():
                 model_name_textbox = gr.Textbox(label="Model name")
                 revision_name_textbox = gr.Textbox(label="revision", placeholder="main")
-                base_model_name_textbox = gr.Textbox(label="base model (for delta)")
+                
             with gr.Column():
-                is_8bit_toggle = gr.Checkbox(False, label="8 bit eval")
-                private = gr.Checkbox(False, label="Private")
+                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")
@@ -220,7 +258,8 @@ with block:
 print("adding refresh leaderboard")
 def refresh_leaderboard():
     leaderboard_table = get_leaderboard()
-    print("leaderboard updated")
+    eval_table = get_eval_table()
+    print("refreshing leaderboard")
 
 scheduler = BackgroundScheduler()
 scheduler.add_job(func=refresh_leaderboard, trigger="interval", seconds=300) # refresh every 5 mins

utils.py

@@ -23,28 +23,13 @@ BENCH_TO_NAME = {
      "hendrycks":"MMLU (5-shot) ⬆️",
      "truthfulqa_mc":"TruthQA (0-shot) ⬆️",
 }
-def make_clickable_model(model_name):
+def make_clickable_model(model_name):        
     # 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>'
 
-def get_n_params(base_model):
-    return "unknown"
-    
-    # WARNING: High memory usage
-
-    # Retrieve the number of parameters from the configuration
-    try:
-        config = AutoConfig.from_pretrained(base_model, use_auth_token=True, low_cpu_mem_usage=True)
-        n_params = AutoModel.from_config(config).num_parameters()
-    except Exception as e:
-        print(f"Error:{e} The number of parameters is not available in the config for the model '{base_model}'.")
-        return "unknown"
-
-    return str(n_params)
-
 @dataclass
 class EvalResult:
     eval_name : str
@@ -66,8 +51,8 @@ class EvalResult:
         data_dict["8bit"] = self.is_8bit
         data_dict["base_model"] = make_clickable_model(base_model)
         data_dict["revision"] = self.revision
-        data_dict["total ⬆️"] = round(sum([v for k,v in self.results.items()]),3)
-        data_dict["# params"] = get_n_params(base_model)
+        data_dict["total ⬆️"] = 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():
@@ -90,7 +75,7 @@ def parse_eval_result(json_filepath: str) -> Tuple[str, dict]:
     model = path_split[-4]
     is_8bit = path_split[-2] == "8bit"
     revision = path_split[-3]
-    if len(path_split)== 6:
+    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:
@@ -101,7 +86,7 @@ def parse_eval_result(json_filepath: str) -> Tuple[str, dict]:
     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),3)
+            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
@@ -109,8 +94,12 @@ def parse_eval_result(json_filepath: str) -> Tuple[str, dict]:
     
     
    
-def get_eval_results() -> List[EvalResult]:
-    json_filepaths = glob.glob("evals/eval_results/**/*.json", recursive=True)
+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:
@@ -125,8 +114,8 @@ def get_eval_results() -> List[EvalResult]:
     
     return eval_results
     
-def get_eval_results_dicts() -> List[Dict]:
-    eval_results = get_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]