Update space

app.py

@@ -1,61 +1,29 @@
 import gradio as gr
-from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
-import pandas as pd
+from gradio_leaderboard import Leaderboard, SelectColumns
 from apscheduler.schedulers.background import BackgroundScheduler
-from huggingface_hub import snapshot_download
 
 from src.about import (
     CITATION_BUTTON_LABEL,
     CITATION_BUTTON_TEXT,
-    EVALUATION_QUEUE_TEXT,
     INTRODUCTION_TEXT,
-    LLM_BENCHMARKS_TEXT,
     TITLE,
 )
 from src.display.css_html_js import custom_css
 from src.display.utils import (
     BENCHMARK_COLS,
     COLS,
-    EVAL_COLS,
-    EVAL_TYPES,
     AutoEvalColumn,
-    ModelType,
     fields,
-    WeightType,
-    Precision
 )
-from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN
-from src.populate import get_evaluation_queue_df, get_leaderboard_df
-from src.submission.submit import add_new_eval
+from src.envs import API, REPO_ID
+from src.populate import get_leaderboard_df
 
 
 def restart_space():
     API.restart_space(repo_id=REPO_ID)
 
-### Space initialisation
-try:
-    print(EVAL_REQUESTS_PATH)
-    snapshot_download(
-        repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
-    )
-except Exception:
-    restart_space()
-try:
-    print(EVAL_RESULTS_PATH)
-    snapshot_download(
-        repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
-    )
-except Exception:
-    restart_space()
-
 
-LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
-
-(
-    finished_eval_queue_df,
-    running_eval_queue_df,
-    pending_eval_queue_df,
-) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
+LEADERBOARD_DF = get_leaderboard_df(COLS, BENCHMARK_COLS)
 
 def init_leaderboard(dataframe):
     if dataframe is None or dataframe.empty:
@@ -68,23 +36,8 @@ def init_leaderboard(dataframe):
             cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden],
             label="Select Columns to Display:",
         ),
-        search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
+        search_columns=[AutoEvalColumn.model.name],
         hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],
-        filter_columns=[
-            ColumnFilter(AutoEvalColumn.model_type.name, type="checkboxgroup", label="Model types"),
-            ColumnFilter(AutoEvalColumn.precision.name, type="checkboxgroup", label="Precision"),
-            ColumnFilter(
-                AutoEvalColumn.params.name,
-                type="slider",
-                min=0.01,
-                max=150,
-                label="Select the number of parameters (B)",
-            ),
-            ColumnFilter(
-                AutoEvalColumn.still_on_hub.name, type="boolean", label="Deleted/incomplete", default=True
-            ),
-        ],
-        bool_checkboxgroup_label="Hide models",
         interactive=False,
     )
 
@@ -95,99 +48,9 @@ with demo:
     gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
 
     with gr.Tabs(elem_classes="tab-buttons") as tabs:
-        with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
+        with gr.TabItem("🎭 RoleRMBench", elem_id="llm-benchmark-tab-table", id=0):
             leaderboard = init_leaderboard(LEADERBOARD_DF)
 
-        with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
-            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
-
-        with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-table", id=3):
-            with gr.Column():
-                with gr.Row():
-                    gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
-
-                with gr.Column():
-                    with gr.Accordion(
-                        f"✅ Finished Evaluations ({len(finished_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            finished_eval_table = gr.components.Dataframe(
-                                value=finished_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-                    with gr.Accordion(
-                        f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            running_eval_table = gr.components.Dataframe(
-                                value=running_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-
-                    with gr.Accordion(
-                        f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})",
-                        open=False,
-                    ):
-                        with gr.Row():
-                            pending_eval_table = gr.components.Dataframe(
-                                value=pending_eval_queue_df,
-                                headers=EVAL_COLS,
-                                datatype=EVAL_TYPES,
-                                row_count=5,
-                            )
-            with gr.Row():
-                gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text")
-
-            with gr.Row():
-                with gr.Column():
-                    model_name_textbox = gr.Textbox(label="Model name")
-                    revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
-                    model_type = gr.Dropdown(
-                        choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],
-                        label="Model type",
-                        multiselect=False,
-                        value=None,
-                        interactive=True,
-                    )
-
-                with gr.Column():
-                    precision = gr.Dropdown(
-                        choices=[i.value.name for i in Precision if i != Precision.Unknown],
-                        label="Precision",
-                        multiselect=False,
-                        value="float16",
-                        interactive=True,
-                    )
-                    weight_type = gr.Dropdown(
-                        choices=[i.value.name for i in WeightType],
-                        label="Weights type",
-                        multiselect=False,
-                        value="Original",
-                        interactive=True,
-                    )
-                    base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")
-
-            submit_button = gr.Button("Submit Eval")
-            submission_result = gr.Markdown()
-            submit_button.click(
-                add_new_eval,
-                [
-                    model_name_textbox,
-                    base_model_name_textbox,
-                    revision_name_textbox,
-                    precision,
-                    weight_type,
-                    model_type,
-                ],
-                submission_result,
-            )
-
     with gr.Row():
         with gr.Accordion("📙 Citation", open=False):
             citation_button = gr.Textbox(
@@ -201,4 +64,4 @@ with demo:
 scheduler = BackgroundScheduler()
 scheduler.add_job(restart_space, "interval", seconds=1800)
 scheduler.start()
-demo.queue(default_concurrency_limit=40).launch()
+demo.queue(default_concurrency_limit=40).launch()

src/about.py

@@ -11,60 +11,25 @@ class Task:
 # Select your tasks here
 # ---------------------------------------------------
 class Tasks(Enum):
-    # task_key in the json file, metric_key in the json file, name to display in the leaderboard 
-    task0 = Task("anli_r1", "acc", "ANLI")
-    task1 = Task("logiqa", "acc_norm", "LogiQA")
+    # task_key, metric_key, name to display in the leaderboard 
+    task0 = Task("avg", "score", "Avg")
+    task1 = Task("nar", "score", "Nar")
+    task2 = Task("mt", "score", "MT")
+    task3 = Task("con", "score", "Con")
+    task4 = Task("if", "score", "IF")
+    task5 = Task("scn", "score", "Scn")
+    task6 = Task("saf", "score", "Saf")
+    task7 = Task("att", "score", "Att")
 
-NUM_FEWSHOT = 0 # Change with your few shot
 # ---------------------------------------------------
 
 
-
 # Your leaderboard name
-TITLE = """<h1 align="center" id="space-title">Demo leaderboard</h1>"""
+TITLE = """<h1 align="center" id="space-title">🎭 RoleRMBench Leaderboard</h1>"""
 
 # What does your leaderboard evaluate?
 INTRODUCTION_TEXT = """
-Intro text
-"""
-
-# Which evaluations are you running? how can people reproduce what you have?
-LLM_BENCHMARKS_TEXT = f"""
-## How it works
-
-## Reproducibility
-To reproduce our results, here is the commands you can run:
-
-"""
-
-EVALUATION_QUEUE_TEXT = """
-## Some good practices before submitting a model
-
-### 1) Make sure you can load your model and tokenizer using AutoClasses:
-```python
-from transformers import AutoConfig, AutoModel, AutoTokenizer
-config = AutoConfig.from_pretrained("your model name", revision=revision)
-model = AutoModel.from_pretrained("your model name", revision=revision)
-tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
-```
-If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded.
-
-Note: make sure your model is public!
-Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!
-
-### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
-It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!
-
-### 3) Make sure your model has an open license!
-This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗
-
-### 4) Fill up your model card
-When we add extra information about models to the leaderboard, it will be automatically taken from the model card
-
-## In case of model failure
-If your model is displayed in the `FAILED` category, its execution stopped.
-Make sure you have followed the above steps first.
-If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task).
+RoleRMBench evaluates reward models on role-playing scenarios across multiple dimensions.
 """
 
 CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"

src/display/utils.py

@@ -1,8 +1,6 @@
 from dataclasses import dataclass, make_dataclass
 from enum import Enum
 
-import pandas as pd
-
 from src.about import Tasks
 
 def fields(raw_class):
@@ -23,88 +21,15 @@ class ColumnContent:
 ## Leaderboard columns
 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)])
+# Scores
 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(["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)])
 
 # We use make dataclass to dynamically fill the scores from Tasks
 AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
 
-## For the queue columns in the submission tab
-@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)
-
-## All the model information that we might need
-@dataclass
-class ModelDetails:
-    name: str
-    display_name: str = ""
-    symbol: str = "" # emoji
-
-
-class ModelType(Enum):
-    PT = ModelDetails(name="pretrained", symbol="🟢")
-    FT = ModelDetails(name="fine-tuned", symbol="🔶")
-    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")
-    RL = ModelDetails(name="RL-tuned", 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 "pretrained" in type or "🟢" in type:
-            return ModelType.PT
-        if "RL-tuned" in type or "🟦" in type:
-            return ModelType.RL
-        if "instruction-tuned" in type or "⭕" in type:
-            return ModelType.IFT
-        return ModelType.Unknown
-
-class WeightType(Enum):
-    Adapter = ModelDetails("Adapter")
-    Original = ModelDetails("Original")
-    Delta = ModelDetails("Delta")
-
-class Precision(Enum):
-    float16 = ModelDetails("float16")
-    bfloat16 = ModelDetails("bfloat16")
-    Unknown = ModelDetails("?")
-
-    def from_str(precision):
-        if precision in ["torch.float16", "float16"]:
-            return Precision.float16
-        if precision in ["torch.bfloat16", "bfloat16"]:
-            return Precision.bfloat16
-        return Precision.Unknown
-
 # Column selection
 COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
 
-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]
-

src/envs.py

@@ -6,20 +6,9 @@ from huggingface_hub import HfApi
 # ----------------------------------
 TOKEN = os.environ.get("HF_TOKEN") # A read/write token for your org
 
-OWNER = "demo-leaderboard-backend" # Change to your org - don't forget to create a results and request dataset, with the correct format!
+OWNER = "demo-leaderboard-backend" # Change to your org
 # ----------------------------------
 
 REPO_ID = f"{OWNER}/leaderboard"
-QUEUE_REPO = f"{OWNER}/requests"
-RESULTS_REPO = f"{OWNER}/results"
-
-# If you setup a cache later, just change HF_HOME
-CACHE_PATH=os.getenv("HF_HOME", ".")
-
-# Local caches
-EVAL_REQUESTS_PATH = os.path.join(CACHE_PATH, "eval-queue")
-EVAL_RESULTS_PATH = os.path.join(CACHE_PATH, "eval-results")
-EVAL_REQUESTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-queue-bk")
-EVAL_RESULTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-results-bk")
 
 API = HfApi(token=TOKEN)

src/leaderboard/read_evals.py

@@ -1,196 +0,0 @@
-import glob
-import json
-import math
-import os
-from dataclasses import dataclass
-
-import dateutil
-import numpy as np
-
-from src.display.formatting import make_clickable_model
-from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
-from src.submission.check_validity import is_model_on_hub
-
-
-@dataclass
-class EvalResult:
-    """Represents one full evaluation. Built from a combination of the result and request file for a given run.
-    """
-    eval_name: str # org_model_precision (uid)
-    full_model: str # org/model (path on hub)
-    org: str 
-    model: str
-    revision: str # commit hash, "" if main
-    results: dict
-    precision: Precision = Precision.Unknown
-    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
-    weight_type: WeightType = WeightType.Original # Original or Adapter
-    architecture: str = "Unknown" 
-    license: str = "?"
-    likes: int = 0
-    num_params: int = 0
-    date: str = "" # submission date of request file
-    still_on_hub: bool = False
-
-    @classmethod
-    def init_from_json_file(self, json_filepath):
-        """Inits the result from the specific model result file"""
-        with open(json_filepath) as fp:
-            data = json.load(fp)
-
-        config = data.get("config")
-
-        # Precision
-        precision = Precision.from_str(config.get("model_dtype"))
-
-        # Get model and org
-        org_and_model = config.get("model_name", config.get("model_args", None))
-        org_and_model = org_and_model.split("/", 1)
-
-        if len(org_and_model) == 1:
-            org = None
-            model = org_and_model[0]
-            result_key = f"{model}_{precision.value.name}"
-        else:
-            org = org_and_model[0]
-            model = org_and_model[1]
-            result_key = f"{org}_{model}_{precision.value.name}"
-        full_model = "/".join(org_and_model)
-
-        still_on_hub, _, model_config = is_model_on_hub(
-            full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False
-        )
-        architecture = "?"
-        if model_config is not None:
-            architectures = getattr(model_config, "architectures", None)
-            if architectures:
-                architecture = ";".join(architectures)
-
-        # Extract results available in this file (some results are split in several files)
-        results = {}
-        for task in Tasks:
-            task = task.value
-
-            # We average all scores of a given metric (not all metrics are present in all files)
-            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])
-            if accs.size == 0 or any([acc is None for acc in accs]):
-                continue
-
-            mean_acc = np.mean(accs) * 100.0
-            results[task.benchmark] = mean_acc
-
-        return self(
-            eval_name=result_key,
-            full_model=full_model,
-            org=org,
-            model=model,
-            results=results,
-            precision=precision,  
-            revision= config.get("model_sha", ""),
-            still_on_hub=still_on_hub,
-            architecture=architecture
-        )
-
-    def update_with_request_file(self, requests_path):
-        """Finds the relevant request file for the current model and updates info with it"""
-        request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)
-
-        try:
-            with open(request_file, "r") as f:
-                request = json.load(f)
-            self.model_type = ModelType.from_str(request.get("model_type", ""))
-            self.weight_type = WeightType[request.get("weight_type", "Original")]
-            self.license = request.get("license", "?")
-            self.likes = request.get("likes", 0)
-            self.num_params = request.get("params", 0)
-            self.date = request.get("submitted_time", "")
-        except Exception:
-            print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")
-
-    def to_dict(self):
-        """Converts the Eval Result to a dict compatible with our dataframe display"""
-        average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
-        data_dict = {
-            "eval_name": self.eval_name,  # not a column, just a save name,
-            AutoEvalColumn.precision.name: self.precision.value.name,
-            AutoEvalColumn.model_type.name: self.model_type.value.name,
-            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
-            AutoEvalColumn.weight_type.name: self.weight_type.value.name,
-            AutoEvalColumn.architecture.name: self.architecture,
-            AutoEvalColumn.model.name: make_clickable_model(self.full_model),
-            AutoEvalColumn.revision.name: self.revision,
-            AutoEvalColumn.average.name: average,
-            AutoEvalColumn.license.name: self.license,
-            AutoEvalColumn.likes.name: self.likes,
-            AutoEvalColumn.params.name: self.num_params,
-            AutoEvalColumn.still_on_hub.name: self.still_on_hub,
-        }
-
-        for task in Tasks:
-            data_dict[task.value.col_name] = self.results[task.value.benchmark]
-
-        return data_dict
-
-
-def get_request_file_for_model(requests_path, model_name, precision):
-    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
-    request_files = os.path.join(
-        requests_path,
-        f"{model_name}_eval_request_*.json",
-    )
-    request_files = glob.glob(request_files)
-
-    # Select correct request file (precision)
-    request_file = ""
-    request_files = sorted(request_files, reverse=True)
-    for tmp_request_file in request_files:
-        with open(tmp_request_file, "r") as f:
-            req_content = json.load(f)
-            if (
-                req_content["status"] in ["FINISHED"]
-                and req_content["precision"] == precision.split(".")[-1]
-            ):
-                request_file = tmp_request_file
-    return request_file
-
-
-def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
-    """From the path of the results folder root, extract all needed info for results"""
-    model_result_filepaths = []
-
-    for root, _, files in os.walk(results_path):
-        # We should only have json files in model results
-        if len(files) == 0 or any([not f.endswith(".json") for f in files]):
-            continue
-
-        # Sort the files by date
-        try:
-            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
-        except dateutil.parser._parser.ParserError:
-            files = [files[-1]]
-
-        for file in files:
-            model_result_filepaths.append(os.path.join(root, file))
-
-    eval_results = {}
-    for model_result_filepath in model_result_filepaths:
-        # Creation of result
-        eval_result = EvalResult.init_from_json_file(model_result_filepath)
-        eval_result.update_with_request_file(requests_path)
-
-        # Store results of same eval together
-        eval_name = eval_result.eval_name
-        if eval_name in eval_results.keys():
-            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
-        else:
-            eval_results[eval_name] = eval_result
-
-    results = []
-    for v in eval_results.values():
-        try:
-            v.to_dict() # we test if the dict version is complete
-            results.append(v)
-        except KeyError:  # not all eval values present
-            continue
-
-    return results

src/populate.py

@@ -1,58 +1,50 @@
-import json
-import os
-
 import pandas as pd
 
-from src.display.formatting import has_no_nan_values, make_clickable_model
-from src.display.utils import AutoEvalColumn, EvalQueueColumn
-from src.leaderboard.read_evals import get_raw_eval_results
-
-
-def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
-    """Creates a dataframe from all the individual experiment results"""
-    raw_data = get_raw_eval_results(results_path, requests_path)
-    all_data_json = [v.to_dict() for v in raw_data]
-
-    df = pd.DataFrame.from_records(all_data_json)
-    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
+from src.display.formatting import make_clickable_model
+
+
+# Static benchmark data
+LEADERBOARD_DATA = [
+    # Open-source Models
+    {"model": "internlm/internlm2-20b-reward", "Avg": 70.58, "Nar": 70.37, "MT": 68.25, "Con": 67.61, "IF": 76.00, "Scn": 72.73, "Saf": 66.10, "Att": 75.00},
+    {"model": "allenai/Llama-3.1-Tulu-3-70B-SFT-RM-RB2", "Avg": 70.36, "Nar": 66.67, "MT": 71.43, "Con": 70.42, "IF": 70.00, "Scn": 65.15, "Saf": 76.27, "Att": 70.59},
+    {"model": "Skywork/Skywork-Reward-V2-Qwen3-8B", "Avg": 70.07, "Nar": 64.81, "MT": 69.84, "Con": 67.61, "IF": 66.00, "Scn": 75.76, "Saf": 74.58, "Att": 77.94},
+    {"model": "internlm/internlm2-7b-reward", "Avg": 67.72, "Nar": 64.81, "MT": 63.49, "Con": 64.79, "IF": 68.00, "Scn": 72.73, "Saf": 72.88, "Att": 66.18},
+    {"model": "allenai/Llama-3.1-Tulu-3-8B-DPO-RM-RB2", "Avg": 67.53, "Nar": 70.37, "MT": 65.08, "Con": 60.56, "IF": 76.00, "Scn": 71.21, "Saf": 67.80, "Att": 61.76},
+    {"model": "allenai/Llama-3.1-70B-Instruct-RM-RB2", "Avg": 66.39, "Nar": 72.22, "MT": 65.08, "Con": 56.34, "IF": 62.00, "Scn": 65.15, "Saf": 76.27, "Att": 67.65},
+    {"model": "allenai/Llama-3.1-Tulu-3-8B-RL-RM-RB2", "Avg": 66.34, "Nar": 70.37, "MT": 61.90, "Con": 60.56, "IF": 72.00, "Scn": 72.73, "Saf": 69.49, "Att": 60.29},
+    {"model": "allenai/Llama-3.1-8B-Instruct-RM-RB2", "Avg": 65.06, "Nar": 59.26, "MT": 61.94, "Con": 59.15, "IF": 70.00, "Scn": 72.73, "Saf": 71.19, "Att": 61.16},
+    {"model": "allenai/Llama-3.1-Tulu-3-8B-SFT-RM-RB2", "Avg": 64.89, "Nar": 66.67, "MT": 60.32, "Con": 57.75, "IF": 70.00, "Scn": 66.67, "Saf": 66.10, "Att": 64.71},
+    {"model": "Skywork/Skywork-Reward-V2-Llama-3.1-8B", "Avg": 64.17, "Nar": 53.70, "MT": 63.49, "Con": 60.56, "IF": 66.00, "Scn": 71.21, "Saf": 69.49, "Att": 64.71},
+    {"model": "CharacterRM", "Avg": 61.11, "Nar": 59.26, "MT": 65.08, "Con": 56.34, "IF": 72.00, "Scn": 66.67, "Saf": 52.54, "Att": 55.88},
+    {"model": "infly/INF-ORM-Llama3.1-70B", "Avg": 58.51, "Nar": 61.11, "MT": 61.90, "Con": 50.70, "IF": 58.00, "Scn": 56.06, "Saf": 64.41, "Att": 57.35},
+    {"model": "Ray2333/GRM_Llama3.1_8B_rewardmodel-ft", "Avg": 56.50, "Nar": 53.70, "MT": 58.73, "Con": 57.75, "IF": 56.00, "Scn": 56.06, "Saf": 59.32, "Att": 52.94},
+    {"model": "Skywork/Skywork-Reward-Llama-3.1-8B", "Avg": 53.50, "Nar": 48.15, "MT": 50.79, "Con": 50.70, "IF": 58.00, "Scn": 59.09, "Saf": 55.93, "Att": 50.00},
+    {"model": "Skywork/Skywork-Reward-Llama-3.1-8B-v0.2", "Avg": 51.97, "Nar": 42.58, "MT": 50.79, "Con": 45.07, "IF": 60.00, "Scn": 50.06, "Saf": 55.93, "Att": 57.35},
+    {"model": "nicolinho/QRM-Llama3.1-8B-v2", "Avg": 47.42, "Nar": 44.44, "MT": 58.73, "Con": 40.85, "IF": 46.00, "Scn": 50.00, "Saf": 43.37, "Att": 48.53},
+    {"model": "NCSOFT/Llama-3-OffsetBias-RM-8B", "Avg": 47.17, "Nar": 44.44, "MT": 49.21, "Con": 39.44, "IF": 32.00, "Scn": 50.00, "Saf": 69.49, "Att": 45.59},
+    # Proprietary Models
+    {"model": "GPT-5-mini-2025-08-07", "Avg": 69.30, "Nar": 68.52, "MT": 73.02, "Con": 59.86, "IF": 83.00, "Scn": 68.94, "Saf": 70.34, "Att": 65.44},
+    {"model": "GPT-4o-2024-08-06", "Avg": 69.12, "Nar": 66.67, "MT": 66.67, "Con": 66.90, "IF": 71.00, "Scn": 68.18, "Saf": 78.81, "Att": 67.65},
+    {"model": "GPT-5-2025-08-07", "Avg": 67.55, "Nar": 69.44, "MT": 66.67, "Con": 66.20, "IF": 82.00, "Scn": 65.91, "Saf": 60.17, "Att": 62.50},
+    {"model": "Claude-3-7-sonnet-20250219", "Avg": 65.24, "Nar": 68.52, "MT": 62.70, "Con": 65.49, "IF": 75.00, "Scn": 62.88, "Saf": 61.02, "Att": 61.76},
+    # Ours
+    {"model": "RoleRM", "Avg": 88.32, "Nar": 90.74, "MT": 82.54, "Con": 80.28, "IF": 94.00, "Scn": 90.91, "Saf": 91.53, "Att": 88.24},
+]
+
+
+def get_leaderboard_df(cols: list, benchmark_cols: list) -> pd.DataFrame:
+    """Creates a dataframe from the static benchmark data"""
+    records = []
+    for entry in LEADERBOARD_DATA:
+        record = {
+            "Model": make_clickable_model(entry["model"]),
+        }
+        for col in benchmark_cols:
+            record[col] = entry[col]
+        records.append(record)
+    
+    df = pd.DataFrame.from_records(records)
+    df = df.sort_values(by=["Avg"], ascending=False)
     df = df[cols].round(decimals=2)
-
-    # filter out if any of the benchmarks have not been produced
-    df = df[has_no_nan_values(df, benchmark_cols)]
     return df
-
-
-def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
-    """Creates the different dataframes for the evaluation queues requestes"""
-    entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
-    all_evals = []
-
-    for entry in entries:
-        if ".json" in entry:
-            file_path = os.path.join(save_path, entry)
-            with open(file_path) as fp:
-                data = json.load(fp)
-
-            data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
-            data[EvalQueueColumn.revision.name] = data.get("revision", "main")
-
-            all_evals.append(data)
-        elif ".md" not in entry:
-            # this is a folder
-            sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")]
-            for sub_entry in sub_entries:
-                file_path = os.path.join(save_path, entry, sub_entry)
-                with open(file_path) as fp:
-                    data = json.load(fp)
-
-                data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
-                data[EvalQueueColumn.revision.name] = data.get("revision", "main")
-                all_evals.append(data)
-
-    pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
-    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
-    finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
-    df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
-    df_running = pd.DataFrame.from_records(running_list, columns=cols)
-    df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
-    return df_finished[cols], df_running[cols], df_pending[cols]

src/submission/check_validity.py

@@ -1,99 +0,0 @@
-import json
-import os
-import re
-from collections import defaultdict
-from datetime import datetime, timedelta, timezone
-
-import huggingface_hub
-from huggingface_hub import ModelCard
-from huggingface_hub.hf_api import ModelInfo
-from transformers import AutoConfig
-from transformers.models.auto.tokenization_auto import AutoTokenizer
-
-def check_model_card(repo_id: str) -> tuple[bool, str]:
-    """Checks if the model card and license exist and have been filled"""
-    try:
-        card = ModelCard.load(repo_id)
-    except huggingface_hub.utils.EntryNotFoundError:
-        return False, "Please add a model card to your model to explain how you trained/fine-tuned it."
-
-    # Enforce license metadata
-    if card.data.license is None:
-        if not ("license_name" in card.data and "license_link" in card.data):
-            return False, (
-                "License not found. Please add a license to your model card using the `license` metadata or a"
-                " `license_name`/`license_link` pair."
-            )
-
-    # Enforce card content
-    if len(card.text) < 200:
-        return False, "Please add a description to your model card, it is too short."
-
-    return True, ""
-
-def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_remote_code=False, test_tokenizer=False) -> tuple[bool, str]:
-    """Checks if the model model_name is on the hub, and whether it (and its tokenizer) can be loaded with AutoClasses."""
-    try:
-        config = AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
-        if test_tokenizer:
-            try:
-                tk = AutoTokenizer.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
-            except ValueError as e:
-                return (
-                    False,
-                    f"uses a tokenizer which is not in a transformers release: {e}",
-                    None
-                )
-            except Exception as e:
-                return (False, "'s tokenizer cannot be loaded. Is your tokenizer class in a stable transformers release, and correctly configured?", None)
-        return True, None, config
-
-    except ValueError:
-        return (
-            False,
-            "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
-            None
-        )
-
-    except Exception as e:
-        return False, "was not found on hub!", None
-
-
-def get_model_size(model_info: ModelInfo, precision: str):
-    """Gets the model size from the configuration, or the model name if the configuration does not contain the information."""
-    try:
-        model_size = round(model_info.safetensors["total"] / 1e9, 3)
-    except (AttributeError, TypeError):
-        return 0  # Unknown model sizes are indicated as 0, see NUMERIC_INTERVALS in app.py
-
-    size_factor = 8 if (precision == "GPTQ" or "gptq" in model_info.modelId.lower()) else 1
-    model_size = size_factor * model_size
-    return model_size
-
-def get_model_arch(model_info: ModelInfo):
-    """Gets the model architecture from the configuration"""
-    return model_info.config.get("architectures", "Unknown")
-
-def already_submitted_models(requested_models_dir: str) -> set[str]:
-    """Gather a list of already submitted models to avoid duplicates"""
-    depth = 1
-    file_names = []
-    users_to_submission_dates = defaultdict(list)
-
-    for root, _, files in os.walk(requested_models_dir):
-        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
-        if current_depth == depth:
-            for file in files:
-                if not file.endswith(".json"):
-                    continue
-                with open(os.path.join(root, file), "r") as f:
-                    info = json.load(f)
-                    file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")
-
-                    # Select organisation
-                    if info["model"].count("/") == 0 or "submitted_time" not in info:
-                        continue
-                    organisation, _ = info["model"].split("/")
-                    users_to_submission_dates[organisation].append(info["submitted_time"])
-
-    return set(file_names), users_to_submission_dates

src/submission/submit.py

@@ -1,119 +0,0 @@
-import json
-import os
-from datetime import datetime, timezone
-
-from src.display.formatting import styled_error, styled_message, styled_warning
-from src.envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
-from src.submission.check_validity import (
-    already_submitted_models,
-    check_model_card,
-    get_model_size,
-    is_model_on_hub,
-)
-
-REQUESTED_MODELS = None
-USERS_TO_SUBMISSION_DATES = None
-
-def add_new_eval(
-    model: str,
-    base_model: str,
-    revision: str,
-    precision: str,
-    weight_type: str,
-    model_type: str,
-):
-    global REQUESTED_MODELS
-    global USERS_TO_SUBMISSION_DATES
-    if not REQUESTED_MODELS:
-        REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH)
-
-    user_name = ""
-    model_path = model
-    if "/" in model:
-        user_name = model.split("/")[0]
-        model_path = model.split("/")[1]
-
-    precision = precision.split(" ")[0]
-    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
-
-    if model_type is None or model_type == "":
-        return styled_error("Please select a model type.")
-
-    # Does the model actually exist?
-    if revision == "":
-        revision = "main"
-
-    # Is the model on the hub?
-    if weight_type in ["Delta", "Adapter"]:
-        base_model_on_hub, error, _ = is_model_on_hub(model_name=base_model, revision=revision, token=TOKEN, test_tokenizer=True)
-        if not base_model_on_hub:
-            return styled_error(f'Base model "{base_model}" {error}')
-
-    if not weight_type == "Adapter":
-        model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, token=TOKEN, test_tokenizer=True)
-        if not model_on_hub:
-            return styled_error(f'Model "{model}" {error}')
-
-    # Is the model info correctly filled?
-    try:
-        model_info = API.model_info(repo_id=model, revision=revision)
-    except Exception:
-        return styled_error("Could not get your model information. Please fill it up properly.")
-
-    model_size = get_model_size(model_info=model_info, precision=precision)
-
-    # Were the model card and license filled?
-    try:
-        license = model_info.cardData["license"]
-    except Exception:
-        return styled_error("Please select a license for your model")
-
-    modelcard_OK, error_msg = check_model_card(model)
-    if not modelcard_OK:
-        return styled_error(error_msg)
-
-    # Seems good, creating the eval
-    print("Adding new eval")
-
-    eval_entry = {
-        "model": model,
-        "base_model": base_model,
-        "revision": revision,
-        "precision": precision,
-        "weight_type": weight_type,
-        "status": "PENDING",
-        "submitted_time": current_time,
-        "model_type": model_type,
-        "likes": model_info.likes,
-        "params": model_size,
-        "license": license,
-        "private": False,
-    }
-
-    # Check for duplicate submission
-    if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:
-        return styled_warning("This model has been already submitted.")
-
-    print("Creating eval file")
-    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
-    os.makedirs(OUT_DIR, exist_ok=True)
-    out_path = f"{OUT_DIR}/{model_path}_eval_request_False_{precision}_{weight_type}.json"
-
-    with open(out_path, "w") as f:
-        f.write(json.dumps(eval_entry))
-
-    print("Uploading eval file")
-    API.upload_file(
-        path_or_fileobj=out_path,
-        path_in_repo=out_path.split("eval-queue/")[1],
-        repo_id=QUEUE_REPO,
-        repo_type="dataset",
-        commit_message=f"Add {model} to eval queue",
-    )
-
-    # Remove the local file
-    os.remove(out_path)
-
-    return styled_message(
-        "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list."
-    )