add rows for llmbenchmark

app.py

@@ -22,12 +22,18 @@ from src.display.utils import (
     ModelType,
     fields,
     WeightType,
-    Precision
+    Precision,
+    NUMERIC_INTERVALS
 )
 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.tools.collections import update_collections
+from src.tools.plots import (
+    create_metric_plot_obj,
+    create_plot_df,
+    create_scores_df,
+)
 
 def restart_space():
     API.restart_space(repo_id=REPO_ID)
@@ -48,45 +54,163 @@ try:
 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)
-
-def init_leaderboard(dataframe):
-    if dataframe is None or dataframe.empty:
-        raise ValueError("Leaderboard DataFrame is empty or None.")
-    return Leaderboard(
-        value=dataframe,
-        datatype=[c.type for c in fields(AutoEvalColumn)],
-        select_columns=SelectColumns(
-            default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
-            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],
-        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,
+    #raw_data, original_df = get_leaderboard_df(
+    leaderboard_df = get_leaderboard_df(
+        results_path=EVAL_RESULTS_PATH,
+        requests_path=EVAL_REQUESTS_PATH,
+        #dynamic_path=DYNAMIC_INFO_FILE_PATH,
+        cols=COLS,
+        benchmark_cols=BENCHMARK_COLS
     )
+    #update_collections(original_df.copy())
+    #leaderboard_df = original_df.copy()
+
+    #plot_df = create_plot_df(create_scores_df(raw_data))
+
+    (
+        finished_eval_queue_df,
+        running_eval_queue_df,
+        pending_eval_queue_df,
+    ) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
+
+    #return leaderboard_df, original_df, plot_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df
+#leaderboard_df, original_df, plot_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = init_space()
+    return leaderboard_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df
+leaderboard_df, finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df = init_space()
+
+
+# Searching and filtering
+def update_table(
+    hidden_df: pd.DataFrame,
+    columns: list,
+    type_query: list,
+    precision_query: str,
+    size_query: list,
+    hide_models: list,
+    query: str,
+):
+    filtered_df = filter_models(df=hidden_df, type_query=type_query, size_query=size_query, precision_query=precision_query, hide_models=hide_models)
+    filtered_df = filter_queries(query, filtered_df)
+    df = select_columns(filtered_df, columns)
+    return df
+
+
+def load_query(request: gr.Request):  # triggered only once at startup => read query parameter if it exists
+    query = request.query_params.get("query") or ""
+    return query, query # return one for the "search_bar", one for a hidden component that triggers a reload only if value has changed
+
+
+def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame:
+    return df[(df[AutoEvalColumn.dummy.name].str.contains(query, case=False))]
+
+
+def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
+    always_here_cols = [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
+    dummy_col = [AutoEvalColumn.dummy.name]
+        #AutoEvalColumn.model_type_symbol.name,
+        #AutoEvalColumn.model.name,
+    # We use COLS to maintain sorting
+    filtered_df = df[
+        always_here_cols + [c for c in COLS if c in df.columns and c in columns] + dummy_col
+    ]
+    return filtered_df
+
+
+def filter_queries(query: str, filtered_df: pd.DataFrame):
+    """Added by Abishek"""
+    final_df = []
+    if query != "":
+        queries = [q.strip() for q in query.split(";")]
+        for _q in queries:
+            _q = _q.strip()
+            if _q != "":
+                temp_filtered_df = search_table(filtered_df, _q)
+                if len(temp_filtered_df) > 0:
+                    final_df.append(temp_filtered_df)
+        if len(final_df) > 0:
+            filtered_df = pd.concat(final_df)
+            filtered_df = filtered_df.drop_duplicates(
+                subset=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name]
+            )
+
+    return filtered_df
+
+
+def filter_models(
+    df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, hide_models: list
+) -> pd.DataFrame:
+    # Show all models
+    if "Private or deleted" in hide_models:
+        filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True]
+    else:
+        filtered_df = df
+
+    if "Contains a merge/moerge" in hide_models:
+        filtered_df = filtered_df[filtered_df[AutoEvalColumn.merged.name] == False]
+
+    if "MoE" in hide_models:
+        filtered_df = filtered_df[filtered_df[AutoEvalColumn.moe.name] == False]
+
+    if "Flagged" in hide_models:
+        filtered_df = filtered_df[filtered_df[AutoEvalColumn.flagged.name] == False]
+
+    type_emoji = [t[0] for t in type_query]
+    filtered_df = filtered_df.loc[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)]
+    filtered_df = filtered_df.loc[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])]
+
+    numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query]))
+    params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors="coerce")
+    mask = params_column.apply(lambda x: any(numeric_interval.contains(x)))
+    filtered_df = filtered_df.loc[mask]
+
+    return filtered_df
+
+leaderboard_df = filter_models(
+    df=leaderboard_df,
+    type_query=[t.to_str(" : ") for t in ModelType],
+    size_query=list(NUMERIC_INTERVALS.keys()),
+    precision_query=[i.value.name for i in Precision],
+    hide_models=[], # Deleted, merges, flagged, MoEs
+)
+
+#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)
+
+#def init_leaderboard(dataframe):
+#    if dataframe is None or dataframe.empty:
+#        raise ValueError("Leaderboard DataFrame is empty or None.")
+#    return Leaderboard(
+#        value=dataframe,
+#        datatype=[c.type for c in fields(AutoEvalColumn)],
+#        select_columns=SelectColumns(
+#            default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
+#            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],
+#        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,
+#    )
 
 
 demo = gr.Blocks(css=custom_css)
@@ -95,8 +219,131 @@ 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):
-            leaderboard = init_leaderboard(LEADERBOARD_DF)
+        with gr.TabItem("🏅 VLM Benchmark", elem_id="vlm-benchmark-tab-table", id=0):
+            #leaderboard = init_leaderboard(LEADERBOARD_DF)
+            with gr.Row():
+                with gr.Column():
+                    with gr.Row():
+                        search_bar = gr.Textbox(
+                            placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...",
+                            show_label=False,
+                            elem_id="search-bar",
+                        )
+                    with gr.Row():
+                        shown_columns = gr.CheckboxGroup(
+                            choices=[
+                                c.name
+                                for c in fields(AutoEvalColumn)
+                                if not c.hidden and not c.never_hidden and not c.dummy
+                            ],
+                            value=[
+                                c.name
+                                for c in fields(AutoEvalColumn)
+                                if c.displayed_by_default and not c.hidden and not c.never_hidden
+                            ],
+                            label="Select columns to show",
+                            elem_id="column-select",
+                            interactive=True,
+                        )
+                    with gr.Row():
+                        hide_models = gr.CheckboxGroup(
+                            label="Hide models",
+                            choices = ["Private or deleted", "Contains a merge/moerge", "Flagged", "MoE"],
+                            value=[],
+                            interactive=True
+                        )
+                with gr.Column(min_width=320):
+                    #with gr.Box(elem_id="box-filter"):
+                    filter_columns_type = gr.CheckboxGroup(
+                        label="Model types",
+                        choices=[t.to_str() for t in ModelType],
+                        value=[t.to_str() for t in ModelType],
+                        interactive=True,
+                        elem_id="filter-columns-type",
+                    )
+                    filter_columns_precision = gr.CheckboxGroup(
+                        label="Precision",
+                        choices=[i.value.name for i in Precision],
+                        value=[i.value.name for i in Precision],
+                        interactive=True,
+                        elem_id="filter-columns-precision",
+                    )
+                    filter_columns_size = gr.CheckboxGroup(
+                        label="Model sizes (in billions of parameters)",
+                        choices=list(NUMERIC_INTERVALS.keys()),
+                        value=list(NUMERIC_INTERVALS.keys()),
+                        interactive=True,
+                        elem_id="filter-columns-size",
+                    )
+
+            leaderboard_table = gr.components.Dataframe(
+                value=leaderboard_df[
+                    [c.name for c in fields(AutoEvalColumn) if c.never_hidden]
+                    + shown_columns.value
+                    + [AutoEvalColumn.dummy.name]
+                ],
+                headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value,
+                datatype=TYPES,
+                elem_id="leaderboard-table",
+                interactive=False,
+                visible=True,
+                #column_widths=["2%", "33%"]
+            )
+
+            # Dummy leaderboard for handling the case when the user uses backspace key
+            hidden_leaderboard_table_for_search = gr.components.Dataframe(
+                value=original_df[COLS],
+                headers=COLS,
+                datatype=TYPES,
+                visible=False,
+            )
+            search_bar.submit(
+                update_table,
+                [
+                    hidden_leaderboard_table_for_search,
+                    shown_columns,
+                    filter_columns_type,
+                    filter_columns_precision,
+                    filter_columns_size,
+                    hide_models,
+                    search_bar,
+                ],
+                leaderboard_table,
+            )
+
+            # Define a hidden component that will trigger a reload only if a query parameter has been set
+            hidden_search_bar = gr.Textbox(value="", visible=False)
+            hidden_search_bar.change(
+                update_table,
+                [
+                    hidden_leaderboard_table_for_search,
+                    shown_columns,
+                    filter_columns_type,
+                    filter_columns_precision,
+                    filter_columns_size,
+                    hide_models,
+                    search_bar,
+                ],
+                leaderboard_table,
+            )
+            # Check query parameter once at startup and update search bar + hidden component
+            demo.load(load_query, inputs=[], outputs=[search_bar, hidden_search_bar])
+
+            for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, hide_models]:
+                selector.change(
+                    update_table,
+                    [
+                        hidden_leaderboard_table_for_search,
+                        shown_columns,
+                        filter_columns_type,
+                        filter_columns_precision,
+                        filter_columns_size,
+                        hide_models,
+                        search_bar,
+                    ],
+                    leaderboard_table,
+                    queue=True,
+                )    
 
         with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
             gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
@@ -201,4 +448,4 @@ 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

@@ -12,8 +12,17 @@ class Task:
 # ---------------------------------------------------
 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")
+    task0 = Task("cmmmu", "acc", "CMMMU")
+    task1 = Task("cmmu", "acc", "CMMU")
+    task2 = Task("cv_bench", "acc", "CV_Bench")
+    task3 = Task("hallusion_bench", "acc", "Hallusion_Bench")
+    task4 = Task("mmmu", "acc", "MMMU")
+    task5 = Task("mmmu_pro_standard", "acc", "MMMU_Pro_Standard")
+    task6 = Task("mmmu_pro_vision", "acc", "MMMU_Pro_Vision")
+    task7 = Task("ocrbench", "acc", "OCRBench")
+    task8 = Task("math_vision", "acc", "Math_Vision")
+    task9 = Task("cvbench", "acc", "CVBench")
+    task10 = Task("ciibench", "acc", "CIIBench")
 
 NUM_FEWSHOT = 0 # Change with your few shot
 # ---------------------------------------------------

src/display/utils.py

@@ -54,6 +54,33 @@ class EvalQueueColumn:  # Queue column
     status = ColumnContent("status", "str", True)
 
 ## All the model information that we might need
+
+# Average ⬆️ human baseline is 0.897 (source: averaging human baselines below)
+# ARC human baseline is 0.80 (source: https://lab42.global/arc/)
+# HellaSwag human baseline is 0.95 (source: https://deepgram.com/learn/hellaswag-llm-benchmark-guide)
+# MMLU human baseline is 0.898 (source: https://openreview.net/forum?id=d7KBjmI3GmQ)
+# TruthfulQA human baseline is 0.94(source: https://arxiv.org/pdf/2109.07958.pdf)
+# Winogrande: https://leaderboard.allenai.org/winogrande/submissions/public
+# GSM8K: paper
+# Define the human baselines
+human_baseline_row = {
+    AutoEvalColumn.model.name: "<p>Human performance</p>",
+    AutoEvalColumn.revision.name: "N/A",
+    AutoEvalColumn.precision.name: None,
+    AutoEvalColumn.average.name: 92.75,
+    AutoEvalColumn.merged.name: False,
+    AutoEvalColumn.arc.name: 80.0,
+    AutoEvalColumn.hellaswag.name: 95.0,
+    AutoEvalColumn.mmlu.name: 89.8,
+    AutoEvalColumn.truthfulqa.name: 94.0,
+    AutoEvalColumn.winogrande.name: 94.0,
+    AutoEvalColumn.gsm8k.name: 100,
+    AutoEvalColumn.c_sem.name: 100,
+    AutoEvalColumn.dummy.name: "human_baseline",
+    AutoEvalColumn.model_type.name: "",
+    AutoEvalColumn.flagged.name: False,
+}
+
 @dataclass
 class ModelDetails:
     name: str
@@ -63,9 +90,9 @@ class ModelDetails:
 
 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="🟦")
+    FT = ModelDetails(name="fine-tuned on domain-specific datasets", symbol="🔶")
+    chat = ModelDetails(name="chat models (RLHF, DPO, IFT, ...)", symbol="💬")
+    merges = ModelDetails(name="base merges and moerges", symbol="🤝")
     Unknown = ModelDetails(name="", symbol="?")
 
     def to_str(self, separator=" "):
@@ -77,10 +104,10 @@ class ModelType(Enum):
             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
+        if any([k in type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "⭕", "💬"]]):
+            return ModelType.chat
+        if "merge" in type or "🤝" in type:
+            return ModelType.merges
         return ModelType.Unknown
 
 class WeightType(Enum):
@@ -91,6 +118,9 @@ class WeightType(Enum):
 class Precision(Enum):
     float16 = ModelDetails("float16")
     bfloat16 = ModelDetails("bfloat16")
+    qt_8bit = ModelDetails("8bit")
+    qt_4bit = ModelDetails("4bit")
+    qt_GPTQ = ModelDetails("GPTQ")
     Unknown = ModelDetails("?")
 
     def from_str(precision):
@@ -98,6 +128,12 @@ class Precision(Enum):
             return Precision.float16
         if precision in ["torch.bfloat16", "bfloat16"]:
             return Precision.bfloat16
+        if precision in ["8bit"]:
+            return Precision.qt_8bit
+        if precision in ["4bit"]:
+            return Precision.qt_4bit
+        if precision in ["GPTQ", "None"]:
+            return Precision.qt_GPTQ
         return Precision.Unknown
 
 # Column selection
@@ -108,3 +144,13 @@ EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]
 
 BENCHMARK_COLS = [t.value.col_name for t in Tasks]
 
+NUMERIC_INTERVALS = {
+    "?": pd.Interval(-1, 0, closed="right"),
+    "~1.5": pd.Interval(0, 2, closed="right"),
+    "~3": pd.Interval(2, 4, closed="right"),
+    "~7": pd.Interval(4, 9, closed="right"),
+    "~13": pd.Interval(9, 20, closed="right"),
+    "~35": pd.Interval(20, 45, closed="right"),
+    "~60": pd.Interval(45, 70, closed="right"),
+    "70+": pd.Interval(70, 10000, closed="right"),
+}

src/populate.py

@@ -6,12 +6,14 @@ 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.leaderboard.filter_models import filter_models_flags
 
 
 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]
+    filter_models_flags(all_data_json)
 
     df = pd.DataFrame.from_records(all_data_json)
     df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)