| import pandas as pd | |
| import gradio as gr | |
| import plotly.express as px | |
| import plotly.graph_objects as go | |
| explanation_data = { | |
| "Accuracy Scores (rename for clarity)": [ | |
| "b4bqa", | |
| "b4b", | |
| "medmcqa_g2b", | |
| "medmcqa_orig_filtered", | |
| "medmcqa_diff", | |
| "medqa_4options_g2b", | |
| "medqa_4options_orig_filtered", | |
| "medqa_diff" | |
| ], | |
| "Description": [ | |
| "Model accuracy on the [Come up with a fitting name] task.", | |
| "[How do we best explain this?]", | |
| "G2B Refers to the 'Generic' to 'Brand' name swap. This is model accuracy on MedMCQA task where generic drug names are substituted with brand names.", | |
| "Model accuracy on MedMCQA task with original data. (Only includes questions that overlap with the g2b dataset)", | |
| "Difference in MedMCQA accuracy for swapped and non-swapped datasets, highlighting the impact of G2B drug name substitution on performance.", | |
| "Model accuracy on MedQA (4 options) task where generic drug names are substituted with brand names.", | |
| "Model accuracy on MedQA (4 options) task with original data. (Only includes questions that overlap with the g2b dataset)", | |
| "Difference in MedMCQA accuracy for swapped and non-swapped datasets, highlighting the impact of G2B drug name substitution on performance." | |
| ] | |
| } | |
| explanation_df = pd.DataFrame(explanation_data) | |
| df = pd.read_csv("data/csv/models_data.csv") | |
| filter_mapping = { | |
| "all": "all", | |
| "🟢 Pre-trained": "🟢", | |
| "🟩 Continuously pre-trained": "🟩", | |
| "🔶 Fine-tuned on domain-specific data": "🔶", | |
| "💬 Chat-models (RLHF, DPO, IFT, ...)": "💬" | |
| } | |
| def filter_items(df, query): | |
| if query == "all": | |
| return df | |
| filter_value = filter_mapping[query] | |
| return df[df["T"].str.contains(filter_value, na=False)] | |
| def create_scatter_plot(df, x_col, y_col, title, x_title, y_title): | |
| fig = px.scatter(df, x=x_col, y=y_col, color='Model', title=title) | |
| fig.add_trace( | |
| go.Scatter( | |
| x=[0, 100], | |
| y=[0, 100], | |
| mode="lines", | |
| name="y=x line", | |
| line=dict(color='black', dash='dash') | |
| ) | |
| ) | |
| fig.update_layout( | |
| xaxis_title=x_title, | |
| yaxis_title=y_title, | |
| xaxis=dict(range=[0, 100]), | |
| yaxis=dict(range=[0, 100]), | |
| legend_title_text='Model' | |
| ) | |
| fig.update_traces(marker=dict(size=10), selector=dict(mode='markers')) | |
| return fig | |
| def create_bar_plot(df, col, title): | |
| sorted_df = df.sort_values(by=col, ascending=True) | |
| fig = px.bar(sorted_df, | |
| x=col, | |
| y='Model', | |
| orientation='h', | |
| title=title, | |
| color=col, | |
| color_continuous_scale='solar') | |
| fig.update_layout(xaxis_title=col, yaxis_title='Model', height=600, coloraxis_showscale=False) | |
| return fig | |
| with gr.Blocks(css="custom.css") as demo: | |
| with gr.Column(): | |
| gr.Markdown( | |
| """<div style="text-align: center;"><h1> <span style='color: #E6B800;'>🐰 RABBITS</span>: <span style='color: #E6B800;'>R</span>obust <span style='color: #E6B800;'>A</span>ssessment of <span style='color: #E6B800;'>B</span>iomedical <span style='color: #E6B800;'>B</span>enchmarks <span style='color: #E6B800;'>I</span>nvolving drug | |
| <span style='color: #E6B800;'>T</span>erm <span style='color: #E6B800;'>S</span>ubstitutions for Language Models <span style='color: #E6B800;'></span></h1></div>""" | |
| ) | |
| with gr.Row(): | |
| gr.Markdown(""" """) | |
| with gr.Row(): | |
| gr.Markdown( | |
| """<div style="text-align: justify;"> | |
| <p class='markdown-text'>Robust language models are crucial in the medical domain and the RABBITS project tests the robustness of LLMs by evaluating their handling of synonyms, specifically brand and generic drug names. We assessed 16 open-source language models from Hugging Face using systematic synonym substitution on MedQA and MedMCQA tasks. Our results show a consistent decline in performance across all model sizes, highlighting challenges in synonym comprehension. Additionally, we discovered significant dataset contamination by identifying overlaps between MedQA, MedMCQA test sets, and the Dolma 1.6 dataset using an 8-gram analysis. This highlights the need to improve model robustness and address contamination in open-source datasets.</p> | |
| </div>""" | |
| ) | |
| with gr.Row(): | |
| gr.Markdown(""" """) | |
| with gr.Row(): | |
| gr.Image(value="workflow-1-2.svg", width=200, height=450) | |
| gr.Image(value="workflow-3-4.svg", width=200, height=450) | |
| with gr.Row(): | |
| gr.Markdown(""" """) | |
| with gr.Row(): | |
| bar1 = gr.Plot( | |
| value=create_bar_plot(df, "medmcqa_diff", "Impact of Generic2Brand swap on MedMCQA Accuracy"), | |
| elem_id="bar1" | |
| ) | |
| bar2 = gr.Plot( | |
| value=create_bar_plot(df, "medqa_diff", "Impact of Generic2Brand swap on MedQA Accuracy"), | |
| elem_id="bar2" | |
| ) | |
| with gr.Row(): | |
| gr.Markdown(""" """) | |
| with gr.Tabs(elem_classes="tab-buttons"): | |
| with gr.TabItem("🔍 Evaluation table"): | |
| with gr.Column(): | |
| with gr.Accordion("➡️ Filter by Column", open=False): | |
| shown_columns = gr.CheckboxGroup( | |
| choices=df.columns.tolist(), | |
| value=df.columns.tolist(), | |
| label="Select Columns", | |
| interactive=True, | |
| ) | |
| with gr.Row(): | |
| search_bar = gr.Textbox( | |
| placeholder="🔍 Search for your model and press ENTER...", | |
| show_label=False, | |
| elem_id="search-bar" | |
| ) | |
| filter_columns = gr.Radio( | |
| label="⏚ Filter model types", | |
| choices=[ | |
| "all", | |
| "🟢 Pre-trained", | |
| "🟩 Continuously pre-trained", | |
| "🔶 Fine-tuned on domain-specific data", | |
| "💬 Chat-models (RLHF, DPO, IFT, ...)" | |
| ], | |
| value="all", | |
| elem_id="filter-columns", | |
| ) | |
| leaderboard_df = gr.Dataframe( | |
| value=df, | |
| headers="keys", | |
| datatype=["html" if col == "Model" else "str" for col in df.columns], | |
| interactive=False, | |
| elem_id="leaderboard-table" | |
| ) | |
| def update_leaderboard(search_query): | |
| filtered_df = df[df["Model"].str.contains(search_query, case=False)] | |
| return filtered_df | |
| search_bar.submit( | |
| update_leaderboard, | |
| inputs=search_bar, | |
| outputs=leaderboard_df | |
| ) | |
| def filter_update(query): | |
| filtered_df = filter_items(df, query) | |
| return filtered_df | |
| filter_columns.change( | |
| filter_update, | |
| inputs=filter_columns, | |
| outputs=leaderboard_df | |
| ) | |
| shown_columns.change( | |
| lambda cols: df[cols], | |
| inputs=shown_columns, | |
| outputs=leaderboard_df | |
| ) | |
| with gr.TabItem("📊 Evaluation Plots"): | |
| with gr.Column(): | |
| with gr.Row(): | |
| scatter1 = gr.Plot( | |
| value=create_scatter_plot(df, "medmcqa_orig_filtered", "medmcqa_g2b", | |
| "MedMCQA: Orig vs G2B", "medmcqa_orig_filtered", "medmcqa_g2b"), | |
| elem_id="scatter1" | |
| ) | |
| scatter2 = gr.Plot( | |
| value=create_scatter_plot(df, "medqa_4options_orig_filtered", "medqa_4options_g2b", | |
| "MedQA: Orig vs G2B", "medqa_4options_orig_filtered", "medqa_4options_g2b"), | |
| elem_id="scatter2" | |
| ) | |
| with gr.Row(): | |
| scatter3 = gr.Plot( | |
| value=create_scatter_plot(df, "b4bqa", "b4b", | |
| "b4bqa vs b4b", "b4bqa", "b4b"), | |
| elem_id="scatter3" | |
| ) | |
| with gr.TabItem("📝 About"): | |
| gr.Markdown( | |
| """<div style="text-align: center;"> | |
| <h2>About RABBITS LLM Leaderboard</h2> | |
| <p>This leaderboard ...</p> | |
| <p>It is designed to ...</p> | |
| </div>""", | |
| elem_classes="markdown-text" | |
| ) | |
| with gr.TabItem("🚀 Submit Here!"): | |
| gr.Markdown( | |
| """<div style="text-align: center;"> | |
| <h2>Submit Your Model Results</h2> | |
| <p>If you have new model results that you would like to add to the leaderboard, please follow the submission guidelines below:</p> | |
| <ul> | |
| <li>COMING SOON</li> | |
| </ul> | |
| <p>COMING SOON</p> | |
| </div>""", | |
| elem_classes="markdown-text" | |
| ) | |
| with gr.Row(): | |
| gr.Dataframe( | |
| value=explanation_df, | |
| headers="keys", | |
| datatype=["str", "str"], | |
| interactive=False, | |
| label="Explanation of Scores" | |
| ) | |
| if __name__ == "__main__": | |
| demo.launch() | |