add RAG experiment tap

app.py

@@ -28,7 +28,7 @@ from src.display.utils import (
     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.populate import get_evaluation_queue_df, get_leaderboard_df, get_rag_leaderboard_df
 from src.submission.submit import add_new_eval
 import base64
 
@@ -122,6 +122,7 @@ except Exception:
 
 # LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
 LEADERBOARD_DF = get_leaderboard_df("leaderboard/data/leaderboard.csv")
+RAG_DF = get_rag_leaderboard_df("leaderboard/data/rag_methods_compare.csv")
 
 # (
 #     finished_eval_queue_df,
@@ -259,10 +260,107 @@ with demo:
 
             gr.Code(code_example_chat, language="shell")
 
-        with gr.TabItem("📝 Details", elem_id="llm-benchmark-tab-table", id=2):
+        
+        with gr.TabItem("🧪 RAG Techniques and Hallucinations", elem_id="llm-benchmark-tab-table", id=2):
+            rag_techniques_markdown = textwrap.dedent(
+                """
+                ## Comparison of Different RAG Techniques and Hallucinations
+
+                Many LLMs can generate fluent answers but still hallucinate facts—especially in RAG settings. This experiment aims to understand how different prompting strategies impact hallucination rates across models. It helps answer: Which prompt format is most reliable? Which models are more sensitive to prompt structure? The goal is to inform better design of RAG pipelines for reducing factual errors in downstream tasks.
+
+                We presents hallucination rates for various LLMs under three different RAG prompting strategies. Each method delivers the same document context and question, but differs in how the information is structured during the prompt.
+
+                ### RAG Techniques Evaluated
+
+                **1. Two-Turn Explicit RAG**  
+                The document and question are sent in separate user messages:
+                ```
+                [System]: You are an assistant for question-answering tasks. 
+                        Given the QUESTION and DOCUMENT you must answer the QUESTION using the information in the DOCUMENT. 
+                        You must not offer new information beyond the context provided in the DOCUMENT. Do not add any external knowledge. 
+                        The ANSWER also must not contradict information provided in the DOCUMENT. 
+                        If the DOCUMENT does not contain the facts to answer the QUESTION or you do not know the answer, you truthfully say that you do not know. 
+                        You have access to information provided by the user as DOCUMENT to answer the QUESTION, and nothing else. 
+                        Use three sentences maximum and keep the answer concise.
+
+                [User]: DOCUMENT: <context>  
+                [User]: QUESTION: <prompt>
+                ```
+                This method creates a multi-turn format, which allows the model to treat the context and question independently.  
+                *Note: This method does not work on Gemma 3 27B due to its restriction on consecutive user messages without an intervening assistant response.*
+
+                **2. System-Prompt Injection RAG**  
+                The document is embedded inside the system prompt, and the user sends only the question:
+                ```
+                [System]: You are an assistant for question-answering tasks. 
+                        Given the QUESTION and DOCUMENT you must answer the QUESTION using the information in the DOCUMENT. 
+                        You must not offer new information beyond the context provided in the DOCUMENT. Do not add any external knowledge. 
+                        The ANSWER also must not contradict information provided in the DOCUMENT. 
+                        If the DOCUMENT does not contain the facts to answer the QUESTION or you do not know the answer, you truthfully say that you do not know. 
+                        You have access to information provided by the user as DOCUMENT to answer the QUESTION, and nothing else. 
+                        Use three sentences maximum and keep the answer concise.
+                        DOCUMENT: <context>  
+
+                [User]: <prompt>
+                ```
+                This approach places the grounding context within the model’s instruction space.
+
+                **3. Single-Turn Concatenated RAG**  
+                Both the document and question are concatenated in a single user message:
+                ```
+                [System]: You are an assistant for question-answering tasks. 
+                        Given the QUESTION and DOCUMENT you must answer the QUESTION using the information in the DOCUMENT. 
+                        You must not offer new information beyond the context provided in the DOCUMENT. Do not add any external knowledge. 
+                        The ANSWER also must not contradict information provided in the DOCUMENT. 
+                        If the DOCUMENT does not contain the facts to answer the QUESTION or you do not know the answer, you truthfully say that you do not know. 
+                        You have access to information provided by the user as DOCUMENT to answer the QUESTION, and nothing else. 
+                        Use three sentences maximum and keep the answer concise.
+                        
+                [User]: 
+                DOCUMENT: <context>  
+                QUESTION: <prompt>
+
+                ```
+                This is the most compact format, sending everything as one prompt input.
+
+                ### Metric
+
+                The values in the table indicate the **hallucination rate (%)** of answers deemed factually incorrect or ungrounded given the provided context.
+
+                """
+
+            
+            )
+
+            gr.Markdown(rag_techniques_markdown, elem_classes="markdown-text")
+
+
+            rag_leaderboard = Leaderboard(
+                value=RAG_DF,
+                datatype=["markdown", "number", "number", "number"],
+                select_columns=SelectColumns(
+                    default_selection=[
+                        "Models",
+                        "Two-Turn Explicit RAG (%)",
+                        "System-Prompt Injection RAG (%)",
+                        "Single-Turn Concatenated RAG (%)"
+                    ],
+                    cant_deselect=["Models"],
+                    label="Select RAG Method Columns:",
+                ),
+                search_columns=["Models"],
+                bool_checkboxgroup_label=None,
+                interactive=False,
+                height=700 
+            )
+
+
+
+
+        with gr.TabItem("📝 Details", elem_id="llm-benchmark-tab-table", id=3):
             gr.Markdown((Path(__file__).parent / "docs.md").read_text(), elem_classes="markdown-text")
 
-        with gr.TabItem("🚀 Submit Here! ", elem_id="llm-benchmark-tab-table", id=3):
+        with gr.TabItem("🚀 Submit Here! ", elem_id="llm-benchmark-tab-table", id=4):
             gr.Markdown((Path(__file__).parent / "submit.md").read_text(), elem_classes="markdown-text")
 
             # with gr.Column():

leaderboard/data/rag_methods_compare.csv

@@ -0,0 +1,13 @@
+Models,rag1,rag2,rag3
+Meta Llama 3.1 8B,11.92,13.09,8.1
+Qwen2.5-VL 7B,13.24,13.39,9.35
+Mistral NeMo,14.42,13.99,10.63
+Meta Llama 4 Maverick,5.72,6.27,3.34
+Meta Llama 4 Scout,6.98,7.17,4.23
+Mistral Small,7.09,7.5,4.74
+Magistral Small,11.87,12.09,8.62
+Gemma 3 27B,-,6.09,3.71
+Meta Llama 3.3 70B,4.65,4.63,2.12
+DeepSeek-V3-0324,7.09,7.71,4.66
+Qwen3-235B-A22B,6.63,6.8,5.04
+DeepSeek-R1-0528,3.58,3.52,2.26

src/populate.py

@@ -4,9 +4,24 @@ 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
+from src.display.utils import EvalQueueColumn
+# from src.leaderboard.read_evals import get_raw_eval_results
 
+def get_rag_leaderboard_df(csv_path):
+    df = pd.read_csv(csv_path)
+
+    for col in ["rag1", "rag2", "rag3"]:
+        df[col] = pd.to_numeric(df[col], errors="coerce").round(2)
+
+    pretty = {
+        "Models":                 "Models",
+        "rag1":     "Two-Turn Explicit RAG (%)",
+        "rag2": "System-Prompt Injection RAG (%)",
+        "rag3": "Single-Turn Concatenated RAG (%)",
+    }
+    df = df.rename(columns=pretty)
+    
+    return df
 
 def get_leaderboard_df(results_path):
     df = pd.read_csv(results_path)