[Add] about

app.py

@@ -1,100 +1,8 @@
 import gradio as gr
 from gradio_leaderboard import Leaderboard, SelectColumns, ColumnFilter
 from pathlib import Path
-import pandas as pd
-
-import os
-
-import json 
-
-import requests
-
-from envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
-from utils import LLM_BENCHMARKS_ABOUT_TEXT, LLM_BENCHMARKS_SUBMIT_TEXT, custom_css, jsonl_to_dataframe, add_average_column_to_df, apply_clickable_model
-
-
-
-def fill_form(model_name, model_id, contact_email, challenge, submission_id, paper_link, architecture, license):
-    value = {
-    # Model name
-    "entry.1591601824": model_name,
-    # username/space
-    "entry.1171388028": model_id,
-    # Submission ID on CMT
-    "entry.171528970": submission_id,
-    # Preprint or paper link
-    "entry.1284338508": paper_link,
-    # Model architecture
-    "entry.1291571256": architecture,
-    # License
-    #   Option: any text
-    "entry.272554778": license,
-    # Challenge
-    #   Option: any text
-    "entry.1908975677": challenge,
-    # Email 
-    #   Option: any text
-    "entry.964644151": contact_email
-    }
-    
-    return value
-    
-def sendForm(url, data):
-    try:
-        requests.post(url, data=data)
-        print("Submitted successfully!")
-    except:
-        print("Error!")
-
-def submit(model_name, model_id, contact_email, challenge, submission_id, paper_link, architecture, license):
-    
-    if model_name == "" or model_id == "" or challenge == "" or architecture == "" or license == "":
-        gr.Error("Please fill all the fields")
-        return
-    if submission_id == "" and paper_link =="":
-        gr.Error("Provide either a link to a paper describing the method or a submission ID for the MLSB workshop.")
-        return
-    try:
-        user_name = ""
-        if "/" in model_id:
-            user_name = model_id.split("/")[0]
-            model_path = model_id.split("/")[1]
-
-        eval_entry = {
-            "model_name": model_name,
-            "model_id": model_id,
-            "challenge": challenge,
-            "submission_id": submission_id,
-            "architecture": architecture,
-            "license": license
-        }
-        OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
-        os.makedirs(OUT_DIR, exist_ok=True)
-        out_path = f"{OUT_DIR}/{user_name}_{model_path}.json"
-
-        with open(out_path, "w") as f:
-            f.write(json.dumps(eval_entry))
-
-        print("Sending form")
-        formData = fill_form(model_name, model_id, contact_email, challenge, submission_id, paper_link, architecture, license)
-        sendForm("https://docs.google.com/forms/d/e/1FAIpQLSf1zP7RAFC5RLlva03xm0eIAPLKXOmMvNUzirbm82kdCUFKNw/formResponse", formData)
-
-        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_name} to eval queue",
-        )
-        
-        gr.Info("Successfully submitted", duration=10)
-        # Remove the local file
-        os.remove(out_path)
-    except:
-        gr.Error("Error submitting the model")
-
 
+from utils import LLM_BENCHMARKS_ABOUT_TEXT, LLM_BENCHMARKS_SUBMIT_TEXT, custom_css, jsonl_to_dataframe, add_average_column_to_df, apply_clickable_model, submit
 
 
 
@@ -180,4 +88,4 @@ with gr.Blocks(css=custom_css) as demo:
         Please find more information about the challenges on [mlsb.io/#challenge](https://mlsb.io/#challenge)""")
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()

utils.py

@@ -1,5 +1,10 @@
-import pandas as pd
 import json
+import os
+
+import gradio as gr
+import pandas as pd
+
+from envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
 
 custom_css = """
 @import url('https://fonts.googleapis.com/css2?family=Vazirmatn&display=swap');
@@ -104,10 +109,46 @@ body, .gradio-container, .gr-button, .gr-input, .gr-slider, .gr-dropdown, .gr-ma
 """
 
 LLM_BENCHMARKS_ABOUT_TEXT = f"""
-## How it works
+## Persian LLM Evaluation Leaderboard (v1)
+
+The Persian LLM Evaluation Leaderboard, developed by **Part DP AI** in collaboration with **AUT (Amirkabir University of Technology) NLP Lab**, provides a comprehensive benchmarking system specifically designed for Persian language models. This leaderboard, based on the open-source [LM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness), offers a unique platform for evaluating the performance of large language models (LLMs) on tasks that demand linguistic proficiency and technical skill in Persian.
+
+## Key Features
+
+1. **Open Evaluation Access**  
+   The leaderboard allows open participation, meaning that developers and researchers working with open-source models can submit evaluation requests for their models. This accessibility encourages the development and testing of Persian LLMs within the broader AI ecosystem.
+
+2. **Task Diversity**  
+   Six specialized tasks have been curated for this leaderboard, each tailored to challenge different aspects of a model’s capabilities. These tasks include:
+   - **Part Multiple Choice**
+   - **ARC Easy**
+   - **ARC Challenging**
+   - **MMLU Pro**
+   - **GSM8k Persian**
+   - **Multiple Choice Persian**
+
+   Each dataset is available in Persian, providing a robust testing ground for models in a non-English setting.
+
+3. **Open-Source Dataset Sample**  
+   A sample of the evaluation dataset is hosted on [Hugging Face Datasets](https://huggingface.co/datasets/PartAI/llm-leaderboard-datasets-sample), offering the AI community a glimpse of the benchmark content and format. This sample allows developers to pre-assess their models against representative data before a full leaderboard evaluation.
 
-## Reproducibility
-To reproduce our results, here is the commands you can run:
+4. **Collaborative Development**  
+   This leaderboard represents a significant collaboration between Part AI and Professor Saeedeh Momtazi of Amirkabir University of Technology, leveraging academic research and industrial expertise to create a high-quality, open benchmarking tool. The partnership underscores a shared commitment to advancing Persian-language AI technologies.
+
+5. **Comprehensive Evaluation Pipeline**  
+   By integrating a standardized evaluation pipeline, models are assessed across a variety of data types, including text, mathematical formulas, and numerical data. This multi-faceted approach enhances the evaluation’s reliability and allows for precise, nuanced assessment of model performance across multiple dimensions.
+
+## Background and Goals
+
+Recent months have seen a notable increase in the development of Persian language models by research centers and AI companies in Iran. However, the lack of reliable, standardized benchmarks for Persian models has made it challenging to evaluate model quality comprehensively. Global benchmarks typically do not support Persian, resulting in skewed or unreliable results for Persian-based AI.
+
+This leaderboard addresses this gap by providing a locally-focused, transparent system that enables consistent, fair comparisons of Persian models. It is expected to be a valuable tool for Persian-speaking businesses and developers, allowing them to select models best suited to their needs. Researchers and model developers also benefit from the competitive environment, with opportunities to showcase and improve their models based on benchmark rankings.
+
+## Data Privacy and Integrity
+
+To maintain evaluation integrity and prevent overfitting or data leakage, only part of the benchmark dataset is openly available. This limited access approach upholds model evaluation reliability, ensuring that results are genuinely representative of each model’s capabilities across unseen data.
+
+The leaderboard represents a significant milestone in Persian language AI and is positioned to become the leading standard for LLM evaluation in the Persian-speaking world.
 
 """
 
@@ -168,4 +209,48 @@ def make_clickable_model(model_name):
 
 def apply_clickable_model(df, column_name):
     df[column_name] = df[column_name].apply(make_clickable_model)
-    return df
+    return df
+
+
+def submit(model_name, model_id, contact_email, challenge, submission_id, paper_link, architecture, license):
+    if model_name == "" or model_id == "" or challenge == "" or architecture == "" or license == "":
+        gr.Error("Please fill all the fields")
+        return
+    if submission_id == "" and paper_link == "":
+        gr.Error("Provide either a link to a paper describing the method or a submission ID for the MLSB workshop.")
+        return
+    try:
+        user_name = ""
+        if "/" in model_id:
+            user_name = model_id.split("/")[0]
+            model_path = model_id.split("/")[1]
+
+        eval_entry = {
+            "model_name": model_name,
+            "model_id": model_id,
+            "challenge": challenge,
+            "submission_id": submission_id,
+            "architecture": architecture,
+            "license": license
+        }
+        OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
+        os.makedirs(OUT_DIR, exist_ok=True)
+        out_path = f"{OUT_DIR}/{user_name}_{model_path}.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_name} to eval queue",
+        )
+
+        gr.Info("Successfully submitted", duration=10)
+        # Remove the local file
+        os.remove(out_path)
+    except:
+        gr.Error("Error submitting the model")