app.py

import gradio as gr
import markdown
import requests

example_dpo_datasets = [
    "mlabonne/orpo-dpo-mix-40k",
    "argilla/ultrafeedback-binarized-preferences-cleaned",
    "argilla/Capybara-Preferences",
]
general_examples = ["davanstrien/cosmochat", "HuggingFaceH4/no_robots"]

datasets_examples = example_dpo_datasets + general_examples


def create_chat_html(messages, dataset_id, offset, compare_mode=False, column=""):
    chat_html = ""
    for turn_number, i in enumerate(range(0, len(messages), 2), start=1):
        user_message = messages[i]
        system_message = messages[i + 1] if i + 1 < len(messages) else None
        user_role = user_message["role"]
        user_content = user_message["content"]
        user_content_html = markdown.markdown(user_content)
        user_content_length = len(user_content)
        user_html = (
            '<div class="user-message" style="justify-content: right;">'
            + '<div class="message-content">'
        )
        user_html += (
            f"<strong>Turn {turn_number} - {user_role.capitalize()}:</strong><br>"
        )
        user_html += f"<em>Length: {user_content_length} characters</em><br><br>"
        user_html += f"{user_content_html}"
        user_html += "</div></div>"
        chat_html += user_html
        if system_message:
            system_role = system_message["role"]
            system_content = system_message["content"]
            system_content_html = markdown.markdown(system_content)
            system_content_length = len(system_content)
            system_html = (
                '<div class="system-message" style="justify-content: left;">'
                + '<div class="message-content">'
            )
            system_html += f"<strong>{system_role.capitalize()}:</strong><br>"
            system_html += (
                f"<em>Length: {system_content_length} characters</em><br><br>"
            )
            system_html += f"{system_content_html}"
            system_html += "</div></div>"
            chat_html += system_html
    if compare_mode:
        chat_html = f'<div class="column {column}">{chat_html}</div>'

    style = """
    <style>
        .user-message, .system-message {
            display: flex;
            margin: 10px;
        }
        .user-message .message-content {
            background-color: #c2e3f7;
            color: #000000;
        }
        .system-message .message-content {
            background-color: #f5f5f5;
            color: #000000;
        }
        .message-content {
            padding: 10px;
            border-radius: 10px;
            max-width: 70%;
            word-wrap: break-word;
        }
        .container {
            display: flex;
            justify-content: space-between;
        }
        .column {
            width: 48%;
        }
    </style>
    """

    dataset_url = f"https://huggingface.co/datasets/{dataset_id}/viewer/default/train?row={offset}"
    dataset_link = f"[View dataset row]({dataset_url})"

    return dataset_link, style + chat_html


def fetch_data(
    dataset_id, chosen_column, rejected_column, current_offset, direction, compare_mode
):
    change = 1 if direction == "Next" else -1
    new_offset = max(0, current_offset + change)

    base_url = f"https://datasets-server.huggingface.co/rows?dataset={dataset_id}&config=default&split=train&offset={new_offset}&length=1"
    response = requests.get(base_url)
    if response.status_code != 200:
        return "", "Failed to fetch data", new_offset
    data = response.json()

    if compare_mode:
        if chosen_column and rejected_column:
            chosen_messages = data["rows"][0]["row"].get(chosen_column, [])
            rejected_messages = data["rows"][0]["row"].get(rejected_column, [])
            chosen_link, chosen_html = create_chat_html(
                chosen_messages,
                dataset_id,
                new_offset,
                compare_mode=True,
                column="chosen",
            )
            rejected_link, rejected_html = create_chat_html(
                rejected_messages,
                dataset_id,
                new_offset,
                compare_mode=True,
                column="rejected",
            )
            chat_html = f'<div class="container">{chosen_html}{rejected_html}</div>'
        else:
            return (
                "",
                "Please provide both chosen and rejected columns for comparison",
                new_offset,
            )
    else:
        if chosen_column:
            messages = data["rows"][0]["row"].get(chosen_column, [])
        else:
            for key, value in data["rows"][0]["row"].items():
                if (
                    isinstance(value, list)
                    and len(value) > 0
                    and isinstance(value[0], dict)
                    and "role" in value[0]
                ):
                    messages = value
                    break
            else:
                return "", "No suitable chat column found", new_offset
        _, chat_html = create_chat_html(messages, dataset_id, new_offset)

    dataset_url = f"https://huggingface.co/datasets/{dataset_id}/viewer/default/train?row={new_offset}"
    dataset_link = f"[View dataset row]({dataset_url})"

    return dataset_link, chat_html, new_offset


def update_column_names(compare_mode):
    return ("chosen", "rejected") if compare_mode else ("", "")


with gr.Blocks() as demo:
    with gr.Row():
        gr.HTML(
            "<h1 style='text-align: center;'>&#x1F4D6; Chat Column Viewer &#x1F4D6;</h1>"
        )
    gr.HTML(
        "<div style='text-align: center;'><em>✨ Explore ChatML formatted data via the datasets viewer API ✨</em></div>"
    )
    gr.Markdown(
        "This app allows you to view chat data from a Hugging Face dataset via the datasets viewer API. ChatML formatted data consists of messages formatted as lists of dictionaries, where each dictionary represents a message with a 'role' (e.g., 'user' or 'assistant') and 'content'. This is a very basic demo built in less than 30 minutes but it hopefully gives you an idea of the kinds of things you can build with the datasets viewer. You can get started building your own apps by going to the datasets viewer documentation [here](https://huggingface.co/docs/datasets-server/index)."
    )
    with gr.Row():
        dataset_id = gr.Dropdown(
            datasets_examples,
            label="Dataset ID",
            allow_custom_value=True,
        )
        chosen_column = gr.Textbox(
            label="Chosen Column",
            placeholder="Column containing chosen chat data",
        )
        rejected_column = gr.Textbox(
            label="Rejected Column",
            placeholder="Column containing rejected chat data",
        )
        compare_mode = gr.Checkbox(label="Compare chosen and rejected chats")
        current_offset = gr.State(value=0)

    with gr.Row():
        back_button = gr.Button("Back")
        next_button = gr.Button("Next")

    dataset_link = gr.Markdown()
    output_html = gr.HTML()

    compare_mode.change(
        fn=update_column_names,
        inputs=compare_mode,
        outputs=[chosen_column, rejected_column],
    )

    back_button.click(
        lambda data, chosen, rejected, offset, compare: fetch_data(
            data, chosen, rejected, offset, "Back", compare
        ),
        inputs=[
            dataset_id,
            chosen_column,
            rejected_column,
            current_offset,
            compare_mode,
        ],
        outputs=[dataset_link, output_html, current_offset],
    )

    next_button.click(
        lambda data, chosen, rejected, offset, compare: fetch_data(
            data, chosen, rejected, offset, "Next", compare
        ),
        inputs=[
            dataset_id,
            chosen_column,
            rejected_column,
            current_offset,
            compare_mode,
        ],
        outputs=[dataset_link, output_html, current_offset],
    )

demo.launch(debug=True)