Spaces:
Running
Running
Commit
·
bad3784
1
Parent(s):
3d11374
added hightlights
Browse files
app.py
CHANGED
@@ -209,16 +209,30 @@ st.write(f"""We evaluate the robustness of the LLM by assessing the variation in
|
|
209 |
# st.write(f"ProbTypos: {other_info_dict['ProbTypos_description']}")
|
210 |
# st.write(f"MaxTypo: {other_info_dict['MaxTypo_description']}")
|
211 |
|
|
|
212 |
for t_opt in options:
|
213 |
st.write('Prompt used : ', t_opt)
|
214 |
t_pert_df_global = result_processor_obj_dict[t_opt].get_global_perturbers_df()
|
|
|
|
|
215 |
|
216 |
t_pert_fig = px.line(t_pert_df_global, x="Levels", y="Performance", color='Perturbation family')
|
217 |
t_pert_fig.update_xaxes(tickmode='linear', dtick=1)
|
218 |
-
|
219 |
|
220 |
st.plotly_chart(t_pert_fig, theme="streamlit", use_container_width=True)
|
221 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
|
223 |
# %%
|
224 |
st.header("Characteristic results")
|
|
|
209 |
# st.write(f"ProbTypos: {other_info_dict['ProbTypos_description']}")
|
210 |
# st.write(f"MaxTypo: {other_info_dict['MaxTypo_description']}")
|
211 |
|
212 |
+
t_pert_df_global_merged = []
|
213 |
for t_opt in options:
|
214 |
st.write('Prompt used : ', t_opt)
|
215 |
t_pert_df_global = result_processor_obj_dict[t_opt].get_global_perturbers_df()
|
216 |
+
t_pert_df_global['Prompt option'] = processed_t_opt_dict[t_opt]
|
217 |
+
t_pert_df_global['Prompt'] = t_pert_df_global['Perturbation family'] + ' - ' + processed_t_opt_dict[t_opt]
|
218 |
|
219 |
t_pert_fig = px.line(t_pert_df_global, x="Levels", y="Performance", color='Perturbation family')
|
220 |
t_pert_fig.update_xaxes(tickmode='linear', dtick=1)
|
221 |
+
t_pert_df_global_merged += [t_pert_df_global]
|
222 |
|
223 |
st.plotly_chart(t_pert_fig, theme="streamlit", use_container_width=True)
|
224 |
|
225 |
+
st.write('---')
|
226 |
+
st.write('The following plot illustrates the variation in performance across perturbation families and prompts.')
|
227 |
+
t_pert_df_global_merged_df = pd.concat(t_pert_df_global_merged, axis=0)
|
228 |
+
t_pert_df_global_merged_df_2 = t_pert_df_global_merged_df[['Prompt option', 'Perturbation family','Levels', 'Performance']]
|
229 |
+
t_pert_df_global_merged_df_2.set_index(t_pert_df_global_merged_df_2.columns[0])
|
230 |
+
# st.dataframe(t_pert_df_global_merged_df_2)
|
231 |
+
|
232 |
+
temp_merged_fig = px.line(t_pert_df_global_merged_df, x="Levels", y="Performance", color='Prompt')
|
233 |
+
temp_merged_fig.update_xaxes(tickmode='linear', dtick=1)
|
234 |
+
st.plotly_chart(temp_merged_fig, theme="streamlit", use_container_width=True)
|
235 |
+
|
236 |
|
237 |
# %%
|
238 |
st.header("Characteristic results")
|