app.py

import gradio as gr
from scipy.stats import anderson_ksamp
from statsmodels.stats.weightstats import ztest
import json
import pandas as pd
from datetime import date
from huggingface_hub import HfApi
from os.path import isfile
from func_timeout import func_timeout, FunctionTimedOut

from scripts.bold import generate_and_evaluate_causal_lm_toxicity_bold
from scripts.honest import calculate_honest, detect_honest_bias
from scripts.winobias import calculate_wino_bias
from avidtools.datamodels.report import Report
from avidtools.datamodels.components import *
from avidtools.datamodels.enums import *

# def generate_report():
def generate_report(classof,type,risk_domain,sep,lifecycle):
# def generate_report(scraped_input, selections):
    report = Report()

    # report.affects = Affects(
    #     developer = [],
    #     deployer = ['Hugging Face'],
    #     artifacts = [Artifact(
    #         type = ArtifactTypeEnum.model,
    #         name = model_id
    #     )]
    # )    
    report.problemtype = Problemtype(
        # classof = clas,
        classof = classof,
        type = type,
        description = LangValue(
            lang = 'eng',
            value = scraped_input['title']
        )
    )
    report.references = [
        Reference(
            label = scraped_input['description'],
            url = scraped_input['url']
        )
    ]
    report.description = LangValue(
        lang = 'eng',
        value = scraped_input['description']
    )
    report.impact = Impact(
        avid = AvidTaxonomy(
            risk_domain = risk_domain,
            sep_view = sep,
            lifecycle_view = lifecycle,
            taxonomy_version = '0.2'
        )
    )
    report.reported_date = date.today()
    
    return report.dict()

scraped_input = {
    "title": "### title",
    "description": "description",
    "url": "https://link.to.arxiv.paper"
}

# selections = {
#     "classof": ClassEnum.llm,
#     "type": TypeEnum.detection,
#     "avid": {
#         "risk_domain": ["Security"],
#         "sep": [SepEnum.E0101],
#         "lifecycle": [LifecycleEnum.L05]
#     }
# }

demo = gr.Blocks(theme=gr.themes.Soft())
# demo = gr.Blocks(theme='gradio/darkdefault')

with demo:

    gr.Markdown("# Report AI Vulnerability Research")
    gr.Markdown("""
    As language models become more prevalent in day-to-day technology, it's important to develop methods to \
    investigate their biases and limitations. To this end, researchers are developing metrics like \
    BOLD, HONEST, and WinoBias that calculate scores which represent their tendency to generate "unfair" text across \
    different collections of prompts. With the widgets below, you can choose a model and a metric to run your own \
    evaluations.
    
    Generating these scores is only half the battle, though! What do you do with these numbers once you've evaluated \
    a model? [AVID](https://avidml.org)'s data model makes it easy to collect and communicate your findings with \
    structured reports.
    """)
    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown("""
            ## Step 1: \n\
            Select a model and a method of detection.
            """)
            with gr.Box():
                title = gr.Markdown(scraped_input['title'])
                description = gr.Markdown(scraped_input['description'])

        with gr.Column(scale=3):
            gr.Markdown("""## Step 2: \
            Categorize your report.""")

            classof = gr.Radio(label="Class", choices=[ce.value for ce in ClassEnum])
            type = gr.Radio(label="Type", choices=[te.value for te in TypeEnum])
            risk_domain = gr.CheckboxGroup(label="Risk Domain", choices=['Security','Ethics','Performance'])
            sep = gr.CheckboxGroup(label="Effect Categories", choices=[se.value for se in SepEnum])
            lifecycle = gr.CheckboxGroup(label="Lifecycle Categories", choices=[le.value for le in LifecycleEnum])

        with gr.Column(scale=5):
            gr.Markdown("""
            ## Step 3: \n\
            Generate a report that you can submit to AVID.

            The title and abstract get auto-populated from Step 1. The taxonomy categories populate from your selections in Step 2.
            """)
            report_button = gr.Button("Generate Report")
            report_json = gr.Json(label="AVID Report")

    report_button.click(
        fn=generate_report,
        inputs=[classof,type,risk_domain,sep,lifecycle],
        outputs=[report_json]
    )

demo.launch()