display plot; use funix 0.6.2

Dockerfile

@@ -6,7 +6,6 @@ COPY ./app/vectara_theme.py /app/vectara_theme.py
 COPY ./app/requirements.txt /app/requirements.txt
 COPY ./app/app.py /app/app.py
 COPY ./app/app_utils.py /app/app_utils.py
-# COPY ./app/results.json /app/results.json
 
 RUN apt-get update && apt-get install -y git-lfs
 

app/app.py

@@ -18,7 +18,7 @@ results_df = load_results()
     direction="column", 
     autorun="always", 
     theme="vectara", 
-    figure_to_image= True,
+    matplotlib_format="svg",
     # output_layout=[
     #     [{"return_index": 0, "width": 0.3}],
     #     [{"return_index": 1, "width": 0.7}],
@@ -26,13 +26,14 @@ results_df = load_results()
 )
 def leaderboard(
     filter_models_by_name: str = ""
-# ) -> Tuple[Markdown, matplotlib.figure.Figure, pd.DataFrame]:
-) -> Tuple[Markdown, pd.DataFrame]:
+    # filter_models_by_name: List[Literal["all", "anthropic", "google", "meta", "openai", "xai", "qwen"]] = ["all"]
+) -> Tuple[Markdown, matplotlib.figure.Figure, pd.DataFrame]:
+# ) -> Tuple[Markdown, pd.DataFrame]:
     """# Hughes Hallucination Evaluation Model (HHEM) Leaderboard
 
     Using [Vectara](https://vectara.com/)'s proprietary [HHEM](https://www.vectara.com/blog/hhem-2-1-a-better-hallucination-detection-model), this leaderboard evaluates how often an LLM hallucinates -- containing information not stated in the source document -- when summarizing a document. For an LLM, its hallucination rate is defined as the ratio of summaries that hallucinate to the total number of summaries it generates. HHEM's open source version is available [here](https://huggingface.co/vectara/hallucination_evaluation_model).    For more details or to contribute, see [this Github repo](https://github.com/vectara/hallucination-leaderboard).
 
-    ## Usage
+    **Usage:**
 
     * All LLMs are displayed by default. To filter, enter the names of the models that you want to see in the "Filter Models by Name" field below, separated by commas or semicolons.
     * Results are paginated. To page thru, use the `<` or `>` buttons at the bottom right corner of the table.
@@ -44,21 +45,18 @@ def leaderboard(
     """
     df = results_df 
 
-    filter_models_by_name = filter_models_by_name.replace(",", ";")
-    filter_models_by_name = filter_models_by_name.replace(" ", "")
-    if len(filter_models_by_name) > 0:
+    filter_models_by_name = filter_models_by_name.replace(",", ";").replace(" ", "")
+    if len(filter_models_by_name) > 0 and "all" not in filter_models_by_name:
         filter_models_by_name = filter_models_by_name.split(";")
+        # filter_models_by_name = [name for name in filter_models_by_name if name != "all"]
         filter_models_by_name = [name for name in filter_models_by_name if name != ""]
         df = df.copy()
-        df = df[df["LLM"].str.contains("|".join(filter_models_by_name), na=False)]
+        df = df[df["LLM_lower_case"].str.contains("|".join(filter_models_by_name), na=False)]
 
         if len(df) == 0: # return an empty DF and an empty figure
-            # return pd.DataFrame(), matplotlib.figure.Figure(), Markdown("No models found")
-            return Markdown("No models found"), pd.DataFrame()
+            return Markdown(f"No models found matching: {filter_models_by_name}"), matplotlib.figure.Figure(), pd.DataFrame()
     
-    return Markdown(""), df
+    # return Markdown(""), df
 
     fig = visualize_leaderboard(df)
-
-    # return df, fig
-    # return Markdown(""), fig, df
+    return Markdown(""), fig, df[["LLM", "Hallucination %", "Answer %", "Avg Summary Words"]]

app/app_utils.py

@@ -5,6 +5,7 @@ from huggingface_hub import Repository
 import pandas as pd
 import matplotlib.pyplot as plt
 import matplotlib.figure
+from datetime import datetime
 from sklearn.preprocessing import MinMaxScaler
 
 # import dotenv
@@ -103,7 +104,7 @@ def load_results(
         print(f"Using pre-dumped results from {results_json}")
 
     results = json.load(open(results_json, "r"))
-    print(results)
+    # print(results)
 
     results_df = pd.DataFrame(results)
     results_df = results_df.sort_values(by="Hallucination %", ascending=True)
@@ -113,6 +114,8 @@ def load_results(
 
     for column in ["Hallucination %", "Answer %", "Avg Summary Words"]:
         results_df[column] = results_df[column].apply(lambda x: round(x, 3))
+
+    results_df["LLM_lower_case"] = results_df["LLM"].str.lower()
     
     return results_df
 
@@ -130,54 +133,57 @@ def determine_font_size(LLM: str, hallucination_percent: float) -> int:
         return 9
     
 def determine_font_color(hallucination_percent: float) -> str:
-    if hallucination_percent < 0.3:
-        return 'white'
-    elif hallucination_percent < 0.65:
+    if 0.25 < hallucination_percent < 0.65:
         return 'black'
     else:
         return 'white'
 
-def determine_llm_x_position(LLM: str, hallucination_percent: float) -> float:
-    # determine the x position of the LLM name
-    # For an LLM, it's bar length is 10* its hallucination %
-    # if the LLM name cannot fit in the bar, move it to the left
-    # if the LLM name can fit in the bar, let its x position be 0.01
-    
+def determine_llm_x_position_and_font_color(LLM: str, hallucination_percent: float) -> float:
     name_length = len(LLM)
     print ("LLM: ", LLM, "hallu_rate: ", hallucination_percent, "name_length: ", name_length)
 
-    hallu_rate_to_bar_length_ratio = 10
+    hallu_rate_to_bar_length_ratio = 5
     bar_length = hallu_rate_to_bar_length_ratio * hallucination_percent
-    if name_length > bar_length:
-        return 0.01
-    else:
-        return hallucination_percent    
-    
+    if name_length < bar_length:
+        return 0.01, determine_font_color(hallucination_percent)
+    else: # to the right of the bar, black anyway
+        return hallucination_percent, 'black'
+
 def visualize_leaderboard(df: pd.DataFrame) -> matplotlib.figure.Figure:
-    fig = plt.figure(figsize=(5, 4))
+    fig = plt.figure(figsize=(8, 4))
     # plot using LLM as x-axis and Hallucination % as y-axis
     # make bars horizontal
     plot_df = df.head(10)
     plot_df["normalized_hallucination_rate"] = min_max_scaler.fit_transform(plot_df[["Hallucination %"]])
 
     plt.barh(plot_df["LLM"], plot_df["Hallucination %"], color=plt.cm.jet(plot_df["normalized_hallucination_rate"]))
+
+    # plot_df["LLM_x_position"], plot_df["font_color"] = zip(*plot_df.apply(
+    #     lambda row: determine_llm_x_position_and_font_color(row["LLM"], row["Hallucination %"]), 
+    #     axis=1
+    # ))
     
-    # for i, row in plot_df.iterrows():
-    #     plt.text(
-    #         determine_llm_x_position(row["LLM"], row["Hallucination %"]), 
-    #         row["LLM"], 
-    #         f"{row['LLM']}", 
-    #         ha='left', 
-    #         va='center', 
-    #         fontsize=9,
-    #         color=determine_font_color(row["normalized_hallucination_rate"])
-    #     )
+    for i, row in plot_df.iterrows():
+        plt.text(
+            # row["LLM_x_position"], 
+            row["Hallucination %"] + 0.025,
+            row["LLM"], 
+            row["Hallucination %"],
+            # f"{row['LLM']}", 
+            ha='left', 
+            va='center', 
+            fontsize=9,
+            # color=row["font_color"]
+        )
     # plt.yticks([])
     plt.tight_layout()
 
+    # add margin to the right of the plot
+    plt.subplots_adjust(right=0.95)
+
     plt.xticks(fontsize=9)
-    # plt.xlabel("Hallucination %", fontsize=9)
-    plt.title("Grounded Hallucination Rate of Best LLMs", fontsize=9)
+    plt.xlabel(f"Copyright (2025) Vectara, Inc. Plot generated on: {datetime.now().strftime('%B %d, %Y')}", fontsize=9)
+    plt.title("Grounded Hallucination Rate of Best LLMs", fontsize=12)
     plt.gca().spines['top'].set_visible(False)
     plt.gca().spines['right'].set_visible(False)
     plt.gca().spines['left'].set_visible(False)

app/requirements.txt

@@ -1,4 +1,4 @@
-funix==0.6.1
+funix==0.6.2
 pandas
 huggingface_hub
 matplotlib