Upload 2 files

app.py

@@ -12,14 +12,16 @@ checkbox_choices = [
     ["Yes", "No", "NA"]
 ]
 
-conversation_data = load_from_jsonl('./data/conversations_unlabeled.jsonl')
-max_conversation_length = max([len(conversation['transcript']) for conversation in conversation_data])
-conversation = get_conversation(conversation_data)
+conversation_data = load_from_jsonl('./data/conversations_unlabeled_sliced.jsonl')
+max_conversation_length = max([len(conversation['transcript']) for conversation in conversation_data_sliced])
 
+conversation = get_conversation(conversation_data_sliced)
 
-def save_labels(conv_id, skipped, submaxim_0=None, submaxim_1=None, submaxim_2=None):
+
+def save_labels(conv_id, slice_idx, skipped, submaxim_0=None, submaxim_1=None, submaxim_2=None):
     data = {
         'conv_id': conv_id,
+        'slice_idx': int(slice_idx),
         'maxim': maxim,
         'skipped': skipped,
         'submaxim_0': submaxim_0,
@@ -28,19 +30,22 @@ def save_labels(conv_id, skipped, submaxim_0=None, submaxim_1=None, submaxim_2=N
     }
     os.makedirs("./labels", exist_ok=True)
 
-    with open(f"./labels/{maxim}_human_labels_{conv_id}.json", 'w') as f:
+    with open(f"./labels/{maxim}_human_labels_{conv_id}_{slice_idx}.json", 'w') as f:
         json.dump(data, f, indent=4)
 
 
 def update_interface(new_conversation):
     new_conv_id = new_conversation['conv_id']
-    new_transcript = pad_transcript(new_conversation['transcript'], max_conversation_length)
+    new_slice_idx = new_conversation['slice_idx']
+    new_transcript = new_conversation['transcript']
+    is_contextual_turn = True if new_slice_idx > 0 else False
 
     markdown_blocks = [None] * max_conversation_length
     for i in range(max_conversation_length):
-        if new_transcript[i]['speaker'] != '':
-            markdown_blocks[i] = gr.Markdown(f"""  **{new_transcript[i]['speaker']}**:      {new_transcript[i]['response']}""",
-                                             visible=True)
+        if i < len(new_transcript) and new_transcript[i]['speaker'] != '':
+            markdown_blocks[i] = gr.Markdown(
+                f"""&nbsp;&nbsp;**{new_transcript[i]['speaker']}**: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;{new_transcript[i]['response']}""",
+                visible=True)
         else:
             markdown_blocks[i] = gr.Markdown("", visible=False)
 
@@ -62,37 +67,40 @@ def update_interface(new_conversation):
     new_radio_2_base = gr.Radio(label=submaxims[2],
                                 choices=checkbox_choices[2],
                                 value=None,
-                                visible=True)
+                                visible=is_contextual_turn)
     conv_len = gr.Number(value=len(new_transcript), visible=False)
 
-    return [new_conv_id] + list(markdown_blocks) + [new_last_response] + [new_radio_0_base] + [new_radio_1_base] + [new_radio_2_base] + [conv_len]
+    return [new_conv_id] + [new_slice_idx] + list(markdown_blocks) + [new_last_response] + [new_radio_0_base] + [new_radio_1_base] + [new_radio_2_base] + [conv_len]
 
 
 def submit(*args):
     conv_id = args[0]
+    slice_idx = args[1]
     submaxim_0 = args[-4]
     submaxim_1 = args[-3]
     submaxim_2 = args[-2]
 
-    save_labels(conv_id, skipped=False, submaxim_0=submaxim_0, submaxim_1=submaxim_1, submaxim_2=submaxim_2)
+    save_labels(conv_id, slice_idx, skipped=False, submaxim_0=submaxim_0, submaxim_1=submaxim_1, submaxim_2=submaxim_2)
 
-    new_conversation = get_conversation(conversation_data)
+    new_conversation = get_conversation(conversation_data_sliced)
     return update_interface(new_conversation)
 
 
 def skip(*args):
     conv_id = args[0]
-    save_labels(conv_id, skipped=True)
+    slice_idx = args[1]
+    save_labels(conv_id, slice_idx, skipped=True)
 
-    new_conversation = get_conversation(conversation_data)
-    return update_interface(new_conversation)
+    new_conversation = get_conversation(conversation_data_sliced)
+    return update_interface(new_conversation, slice_idx)
 
 
 with gr.Blocks(theme=gr.themes.Default()) as interface:
     conv_id = conversation['conv_id']
+    slice_idx = conversation['slice_idx']
     transcript = conversation['transcript']
     conv_len = gr.Number(value=len(transcript), visible=False)
-    padded_transcript = pad_transcript(transcript, max_conversation_length)
+    is_contextual_turn = True if slice_idx > 0 else False
 
     markdown_blocks = [None] * max_conversation_length
     with gr.Column(scale=1, min_width=600):
@@ -100,7 +108,11 @@ with gr.Blocks(theme=gr.themes.Default()) as interface:
             gr.Markdown("""<span style='font-size: 16px;'>&nbsp;&nbsp;&nbsp;&nbsp;**Conversational context** </span>""",
                         visible=True)
         for i in range(max_conversation_length):
-            markdown_blocks[i] = gr.Markdown(f"""&nbsp;&nbsp;**{padded_transcript[i]['speaker']}**: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;{padded_transcript[i]['response']}""")
+            if i < len(transcript):
+                markdown_blocks[i] = gr.Markdown(
+                    f"""&nbsp;&nbsp;**{transcript[i]['speaker']}**: &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;{transcript[i]['response']}""")
+            else:
+                markdown_blocks[i] = gr.Markdown("")
             if i >= conv_len.value:
                 markdown_blocks[i].visible = False
 
@@ -123,17 +135,19 @@ with gr.Blocks(theme=gr.themes.Default()) as interface:
                 radio_submaxim_1_base = gr.Radio(label=submaxims[1],
                                                  choices=checkbox_choices[1],
                                                  value=None,
-                                                 visible=True)
+                                                 visible=is_contextual_turn)
                 radio_submaxim_2_base = gr.Radio(label=submaxims[2],
                                                  choices=checkbox_choices[2],
                                                  value=None,
-                                                 visible=True)
+                                                 visible=is_contextual_turn)
 
     submit_button = gr.Button("Submit")
     skip_button = gr.Button("Skip")
 
     conv_id_element = gr.Text(value=conv_id, visible=False)
+    slice_idx_element = gr.Text(value=slice_idx, visible=False)
     input_list = [conv_id_element] + \
+                 [slice_idx_element] + \
                  markdown_blocks + \
                  [last_response] + \
                  [radio_submaxim_0_base] + \
@@ -144,6 +158,7 @@ with gr.Blocks(theme=gr.themes.Default()) as interface:
         fn=submit,
         inputs=input_list,
         outputs=[conv_id_element,
+                 slice_idx_element,
                  *markdown_blocks,
                  last_response,
                  radio_submaxim_0_base,
@@ -155,6 +170,7 @@ with gr.Blocks(theme=gr.themes.Default()) as interface:
         fn=skip,
         inputs=input_list,
         outputs=[conv_id_element,
+                 slice_idx_element,
                  *markdown_blocks,
                  last_response,
                  radio_submaxim_0_base,
@@ -165,17 +181,15 @@ with gr.Blocks(theme=gr.themes.Default()) as interface:
 
 css = """
 #textbox_id textarea {
-    background-color: white;
+   background-color: white;
 }
 
 .bottom-aligned-group {
-    display: flex;
-    flex-direction: column;
-    justify-content: flex-end;
-    height: 100%;
+   display: flex;
+   flex-direction: column;
+   justify-content: flex-end;
+   height: 100%;
 }
 """
 interface.css = css
 interface.launch()
-
-

interface_utils.py

@@ -31,17 +31,24 @@ def save_to_jsonl(data, filename):
             file.write(json_line + '\n')
 
 
-def get_conversation(conversation_data):
-    conv = random.choice(conversation_data)
-    return conv
-
-
-def pad_transcript(transcript, max_length):
-    padding_count = max_length - len(transcript)
-    if padding_count > 0:
-        for _ in range(padding_count):
-            transcript.append({'speaker': '', 'response': ''})
-    return transcript
+def get_conversation(data, min_length=0):
+    conv = random.choice(data)
+    transcript = conv['transcript']
+    slice_index = random.randint(min_length, len(transcript) - 1)
+    conv_slice = transcript[slice_index]
+    return {
+        'conv_id': conv['conv_id'],
+        'slice_idx': slice_index,
+        'transcript': conv_slice
+    }
+
+
+# def pad_transcript(transcript, max_length):
+#     padding_count = max_length - len(transcript)
+#     if padding_count > 0:
+#         for _ in range(padding_count):
+#             transcript.append({'speaker': '', 'response': ''})
+#     return transcript
 
 
 def get_last_response(transcript):