control handle impossible

app.py

@@ -143,7 +143,7 @@ def init_models():
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     question_answerer = pipeline(
         "question-answering", model='sultan/BioM-ELECTRA-Large-SQuAD2-BioASQ8B',
-        device=device, handle_impossible_answer=True,
+        device=device, handle_impossible_answer=False,
     )
     reranker = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', device=device)
     # queryexp_tokenizer = AutoTokenizer.from_pretrained("doc2query/all-with_prefix-t5-base-v1")
@@ -214,6 +214,9 @@ with st.expander("Settings (strictness, context limit, top hits)"):
     concat_passages = st.radio(
         "Concatenate passages as one long context?",
         ('yes', 'no'))
+    present_impossible = st.radio(
+        "Present impossible answers? (if the model thinks its impossible to answer should it still try?)",
+        ('no', 'yes'))
     support_all = st.radio(
         "Use abstracts and titles as a ranking signal (if the words are matched in the abstract then the document is more relevant)?",
         ('no', 'yes'))
@@ -323,10 +326,10 @@ def run_query(query, progress_bar):
     progress_bar.progress(50)
     if concat_passages == 'yes':
         context = '\n---'.join(contexts)
-        model_results = qa_model(question=query, context=context, top_k=10, doc_stride=512 // 2, max_answer_len=128, max_seq_len=512)
+        model_results = qa_model(question=query, context=context, top_k=10, doc_stride=512 // 2, max_answer_len=128, max_seq_len=512, handle_impossible_answer=present_impossible=='yes')
     else:
         context = ['\n---\n'+ctx for ctx in contexts]
-        model_results = qa_model(question=[query]*len(contexts), context=context)
+        model_results = qa_model(question=[query]*len(contexts), context=context, handle_impossible_answer=present_impossible=='yes')
 
     results = []