app.py

import gradio as gr
import os
from interface_utils import *

maxim = 'quality'
submaxims = ["The response is factual and supported by adequate evidence whenever possible."]
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)


def save_labels(conv_id, skipped, submaxim_0=None):
    data = {
        'conv_id': conv_id,
        'maxim': maxim,
        'skipped': skipped,
        'submaxim_0': submaxim_0
    }
    os.makedirs("./labels", exist_ok=True)

    with open(f"./labels/{maxim}_human_labels_{conv_id}.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)

    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)
        else:
            markdown_blocks[i] = gr.Markdown("", visible=False)

    new_last_response = gr.Text(value=get_last_response(new_transcript),
                                label="",
                                lines=1,
                                container=False,
                                interactive=False,
                                autoscroll=True,
                                visible=True)
    new_radio_0_base = gr.Radio(label=submaxims[0],
                                choices=checkbox_choices[0],
                                value=None,
                                visible=True)
    conv_len = gr.Number(value=len(new_transcript), visible=False)

    return [new_conv_id] + list(markdown_blocks) + [new_last_response] + [new_radio_0_base] + [conv_len]


def submit(*args):
    conv_id = args[0]
    submaxim_0 = args[-2]

    save_labels(conv_id, skipped=False, submaxim_0=submaxim_0)

    new_conversation = get_conversation(conversation_data)
    return update_interface(new_conversation)


def skip(*args):
    conv_id = args[0]
    save_labels(conv_id, skipped=True)

    new_conversation = get_conversation(conversation_data)
    return update_interface(new_conversation)


with gr.Blocks(theme=gr.themes.Default()) as interface:
    conv_id = conversation['conv_id']
    transcript = conversation['transcript']
    conv_len = gr.Number(value=len(transcript), visible=False)
    padded_transcript = pad_transcript(transcript, max_conversation_length)

    markdown_blocks = [None] * max_conversation_length
    with gr.Column(scale=1, min_width=600):
        with gr.Group():
            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 >= conv_len.value:
                markdown_blocks[i].visible = False

        with gr.Row():
            with gr.Group(elem_classes="bottom-aligned-group"):
                speaker_adapted = gr.Markdown(
                    f"""<span style='font-size: 16px;'>&nbsp;&nbsp;&nbsp;&nbsp;**Response to label** </span>""",
                    visible=True)
                last_response = gr.Textbox(value=get_last_response(transcript),
                                           label="",
                                           lines=1,
                                           container=False,
                                           interactive=False,
                                           autoscroll=True,
                                           visible=True)
                radio_submaxim_0_base = gr.Radio(label=submaxims[0],
                                                 choices=checkbox_choices[0],
                                                 value=None,
                                                 visible=True)

    submit_button = gr.Button("Submit")
    skip_button = gr.Button("Skip")

    conv_id_element = gr.Text(value=conv_id, visible=False)
    input_list = [conv_id_element] + \
                 markdown_blocks + \
                 [last_response] + \
                 [radio_submaxim_0_base] + \
                 [conv_len]
    submit_button.click(
        fn=submit,
        inputs=input_list,
        outputs=[conv_id_element,
                 *markdown_blocks,
                 last_response,
                 radio_submaxim_0_base,
                 conv_len]
    )
    skip_button.click(
        fn=skip,
        inputs=input_list,
        outputs=[conv_id_element,
                 *markdown_blocks,
                 last_response,
                 radio_submaxim_0_base,
                 conv_len]
    )

css = """
#textbox_id textarea {
    background-color: white;
}

.bottom-aligned-group {
    display: flex;
    flex-direction: column;
    justify-content: flex-end;
    height: 100%;
}
"""
interface.css = css
interface.launch()