app.py

import requests, time, json, gradio as gr, pandas as pd, ast

def agi_answer(endpoint, instruction, question, temp, top_p, top_k, beams, max_tokens):
    response = requests.post(f'{endpoint}/run/predict', json={
    	'data': [
    		instruction,
    		question,
    		temp,
    		top_p,
    		top_k,
    		beams,
    		max_tokens,
    	]
    }).json()

    return response['data'][0]

def eval_agi(endpoint, temp, top_p, top_k, beams, max_tokens):
    test_df = pd.read_csv('mmlu_testdf.csv')
    for index, value in test_df['input'].iteritems():
        
        ans = agi_answer(endpoint, ast.literal_eval(value)[0]['content'], ast.literal_eval(value)[1]['content'], temp, top_p, top_k, beams, max_tokens)
        test_df.loc[index, 'Answer_AGI'] = ans[:1]
        test_df.loc[index, 'Answer_AGI_raw'] = ans
        print(index, '/', test_df.shape[0])
        time.sleep(0.001)
    accuracy = (test_df['ideal'] == test_df['Answer_AGI']).sum() / len(test_df)
    return [accuracy, test_df[['ideal', 'Answer_AGI']]]

demo = gr.Interface(fn=eval_agi,
                    inputs=[
                        gr.inputs.Textbox(default='https://191779ad955db5c67f.gradio.live', label='endpoint'),
                        gr.inputs.Slider(0, 1, label='temperature', default=0.1),
                        gr.inputs.Slider(0, 1, default=0.75, label='top p'),
                        gr.inputs.Slider(0, 100, default=40, label='top k'),
                        gr.inputs.Slider(0, 4, default=4, label='beams'),
                        gr.inputs.Slider(0, 2000, default=128, label='max tokens')
                    ],
                    outputs=[
                        gr.outputs.Label(label="Accuracy"),
                        'dataframe'
                        ])

demo.launch()