feat-add-reranker-tab-0607 (#21)

- feat: add the ranking only tab for qa (ca1267e31152cb65545c46a9c5d52d015bb8f1f9)
- feat: add the ranking only tab for long-doc (de0db296fa5eb71e078501722b44ef22b61f7e26)
- feat: fix the reranker list (6d20d9786aaf9786d84f0729c935438a0e9ec1fa)
- chore: clean up (118f4b013300f1bb97c990755dee3d7339cbb240)

app.py

@@ -11,10 +11,10 @@ from src.about import (
 from src.benchmarks import DOMAIN_COLS_QA, LANG_COLS_QA, DOMAIN_COLS_LONG_DOC, LANG_COLS_LONG_DOC, METRIC_LIST, \
     DEFAULT_METRIC_QA, DEFAULT_METRIC_LONG_DOC
 from src.display.css_html_js import custom_css
-from src.display.utils import COL_NAME_IS_ANONYMOUS, COL_NAME_REVISION, COL_NAME_TIMESTAMP, COL_NAME_RERANKING_MODEL
+from src.display.utils import COL_NAME_IS_ANONYMOUS, COL_NAME_REVISION, COL_NAME_TIMESTAMP, COL_NAME_RERANKING_MODEL, COL_NAME_RETRIEVAL_MODEL
 from src.envs import API, EVAL_RESULTS_PATH, REPO_ID, RESULTS_REPO, TOKEN
 from src.read_evals import get_raw_eval_results, get_leaderboard_df
-from src.utils import update_metric, upload_file, get_default_cols, submit_results, reset_rank
+from src.utils import update_metric, upload_file, get_default_cols, submit_results, reset_rank, remove_html
 from src.display.gradio_formatting import get_version_dropdown, get_search_bar, get_reranking_dropdown, \
     get_metric_dropdown, get_domain_dropdown, get_language_dropdown, get_anonymous_checkbox, get_revision_and_ts_checkbox, get_leaderboard_table, get_noreranking_dropdown
 from src.display.gradio_listener import set_listeners
@@ -108,9 +108,8 @@ with demo:
                         show_anonymous = get_anonymous_checkbox()
                     with gr.Row():
                         show_revision_and_timestamp = get_revision_and_ts_checkbox()
-
             with gr.Tabs(elem_classes="tab-buttons") as sub_tabs:
-                with gr.TabItem("Retriever + Reranker", id=10):
+                with gr.TabItem("Retrieval + Reranking", id=10):
                     with gr.Row():
                         # search retrieval models
                         with gr.Column():
@@ -149,17 +148,19 @@ with demo:
                         leaderboard_table,
                         queue=True
                     )
-                with gr.TabItem("Retriever Only", id=11):
-                    with gr.Column():
-                        search_bar_retriever = get_search_bar()
-                    selected_noreranker = get_noreranking_dropdown()
+                with gr.TabItem("Retrieval Only", id=11):
+                    with gr.Row():
+                        with gr.Column(scale=1):
+                            search_bar_retriever = get_search_bar()
+                        with gr.Column(scale=1):
+                            selected_noreranker = get_noreranking_dropdown()
                     lb_df_retriever = leaderboard_df_qa[leaderboard_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
                     lb_df_retriever = reset_rank(lb_df_retriever)
-                    hidden_lb_db_retriever = original_df_qa[original_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
-                    hidden_lb_db_retriever = reset_rank(hidden_lb_db_retriever)
                     lb_table_retriever = get_leaderboard_table(lb_df_retriever, types_qa)
                     # Dummy leaderboard for handling the case when the user uses backspace key
-                    hidden_lb_table_retriever = get_leaderboard_table(hidden_lb_db_retriever, types_qa, visible=False)
+                    hidden_lb_df_retriever = original_df_qa[original_df_qa[COL_NAME_RERANKING_MODEL] == "NoReranker"]
+                    hidden_lb_df_retriever = reset_rank(hidden_lb_df_retriever)
+                    hidden_lb_table_retriever = get_leaderboard_table(hidden_lb_df_retriever, types_qa, visible=False)
 
                     set_listeners(
                         "qa",
@@ -188,7 +189,48 @@ with demo:
                         lb_table_retriever,
                         queue=True
                     )
+                with gr.TabItem("Reranking Only", id=12):
+                    lb_df_reranker = leaderboard_df_qa[leaderboard_df_qa[COL_NAME_RETRIEVAL_MODEL] == "BM25"]
+                    lb_df_reranker = reset_rank(lb_df_reranker)
+                    reranking_models_reranker = lb_df_reranker[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
+                    with gr.Row():
+                        with gr.Column(scale=1):
+                            selected_rerankings_reranker = get_reranking_dropdown(reranking_models_reranker)
+                        with gr.Column(scale=1):
+                            search_bar_reranker = gr.Textbox(show_label=False, visible=False)
+                    lb_table_reranker = get_leaderboard_table(lb_df_reranker, types_qa)
+                    hidden_lb_df_reranker = original_df_qa[original_df_qa[COL_NAME_RETRIEVAL_MODEL] == "BM25"]
+                    hidden_lb_df_reranker = reset_rank(hidden_lb_df_reranker)
+                    hidden_lb_table_reranker = get_leaderboard_table(
+                        hidden_lb_df_reranker, types_qa, visible=False
+                    )
 
+                    set_listeners(
+                        "qa",
+                        lb_table_reranker,
+                        hidden_lb_table_reranker,
+                        search_bar_reranker,
+                        selected_domains,
+                        selected_langs,
+                        selected_rerankings_reranker,
+                        show_anonymous,
+                        show_revision_and_timestamp,
+                    )
+                    # set metric listener
+                    selected_metric.change(
+                        update_metric_qa,
+                        [
+                            selected_metric,
+                            selected_domains,
+                            selected_langs,
+                            selected_rerankings_reranker,
+                            search_bar_reranker,
+                            show_anonymous,
+                            show_revision_and_timestamp,
+                        ],
+                        lb_table_reranker,
+                        queue=True
+                    )
         with gr.TabItem("Long Doc", elem_id="long-doc-benchmark-tab-table", id=1):
             with gr.Row():
                 with gr.Column(min_width=320):
@@ -211,7 +253,7 @@ with demo:
                     with gr.Row():
                         show_revision_and_timestamp = get_revision_and_ts_checkbox()
             with gr.Tabs(elem_classes="tab-buttons") as sub_tabs:
-                with gr.TabItem("Retriever + Reranker", id=20):
+                with gr.TabItem("Retrieval + Reranking", id=20):
                     with gr.Row():
                         with gr.Column():
                             search_bar = get_search_bar()
@@ -255,11 +297,12 @@ with demo:
                         lb_table,
                         queue=True
                     )
-
-                with gr.TabItem("Retriever Only", id=21):
-                    with gr.Column():
-                        search_bar_retriever = get_search_bar()
-                    selected_noreranker = get_noreranking_dropdown()
+                with gr.TabItem("Retrieval Only", id=21):
+                    with gr.Row():
+                        with gr.Column(scale=1):
+                            search_bar_retriever = get_search_bar()
+                        with gr.Column(scale=1):
+                            selected_noreranker = get_noreranking_dropdown()
                     lb_df_retriever_long_doc = leaderboard_df_long_doc[
                         leaderboard_df_long_doc[COL_NAME_RERANKING_MODEL] == "NoReranker"
                     ]
@@ -300,6 +343,49 @@ with demo:
                         lb_table_retriever_long_doc,
                         queue=True
                     )
+                with gr.TabItem("Reranking Only", id=22):
+                    lb_df_reranker_ldoc = leaderboard_df_long_doc[
+                        leaderboard_df_long_doc[COL_NAME_RETRIEVAL_MODEL] == "BM25"
+                        ]
+                    lb_df_reranker_ldoc = reset_rank(lb_df_reranker_ldoc)
+                    reranking_models_reranker_ldoc = lb_df_reranker_ldoc[COL_NAME_RERANKING_MODEL].apply(remove_html).unique().tolist()
+                    with gr.Row():
+                        with gr.Column(scale=1):
+                            selected_rerankings_reranker_ldoc = get_reranking_dropdown(reranking_models_reranker_ldoc)
+                        with gr.Column(scale=1):
+                            search_bar_reranker_ldoc = gr.Textbox(show_label=False, visible=False)
+                    lb_table_reranker_ldoc = get_leaderboard_table(lb_df_reranker_ldoc, types_long_doc)
+                    hidden_lb_df_reranker_ldoc = original_df_long_doc[original_df_long_doc[COL_NAME_RETRIEVAL_MODEL] == "BM25"]
+                    hidden_lb_df_reranker_ldoc = reset_rank(hidden_lb_df_reranker_ldoc)
+                    hidden_lb_table_reranker_ldoc = get_leaderboard_table(
+                        hidden_lb_df_reranker_ldoc, types_long_doc, visible=False
+                    )
+
+                    set_listeners(
+                        "long-doc",
+                        lb_table_reranker_ldoc,
+                        hidden_lb_table_reranker_ldoc,
+                        search_bar_reranker_ldoc,
+                        selected_domains,
+                        selected_langs,
+                        selected_rerankings_reranker_ldoc,
+                        show_anonymous,
+                        show_revision_and_timestamp,
+                    )
+                    selected_metric.change(
+                        update_metric_long_doc,
+                        [
+                            selected_metric,
+                            selected_domains,
+                            selected_langs,
+                            selected_rerankings_reranker_ldoc,
+                            search_bar_reranker_ldoc,
+                            show_anonymous,
+                            show_revision_and_timestamp,
+                        ],
+                        lb_table_reranker_ldoc,
+                        queue=True
+                    )
 
         with gr.TabItem("🚀Submit here!", elem_id="submit-tab-table", id=2):
             with gr.Column():