danlou's picture
include example code
03d5009
metadata
license: llama2

The code below shows how this Buyer Persona generator can be used.

More documentation coming soon...

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from tqdm import tqdm

device = "cuda" if torch.cuda.is_available() else "cpu"

model_id = "danlou/persona-generator-llama-2-7b-qlora-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.float16)


def chunks(lst, n):
    """Yield successive n-sized chunks from lst."""
    for i in range(0, len(lst), n):
        yield lst[i:i + n]


def parse_outputs(output_text):

    try:
        output_lns = output_text.split('\n')
        assert len(output_lns) == 2
        assert len(output_lns[0].split(',')) == 2
        assert len(output_lns[1]) > 16

        name, age = [s.strip() for s in output_lns[0].split(',')]
        desc = output_lns[1].strip()

    except AssertionError:
        raise Exception('Malformed output.')
    
    try:
        age = int(age)
    except ValueError:
        raise Exception('Malformed output (age).')
    
    return {'name': name, 'age': age, 'description': desc}



def generate_personas(product, n=1, batch_size=32, parse=True):

    prompt = f"### Instruction:\nDescribe the ideal persona for this product:\n{product}\n\n### Response:\n"
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)

    personas = []
    with tqdm(total=n) as pbar:
        for batch in chunks(range(n), batch_size):
            outputs = model.generate(input_ids,
                                    do_sample=True,
                                    num_beams=1,
                                    num_return_sequences=len(batch),
                                    max_length=512,
                                    min_length=32,
                                    temperature=0.9)

            for output_ids in outputs:
                output_decoded = tokenizer.decode(output_ids, skip_special_tokens=True)
                output_decoded = output_decoded[len(prompt):].strip()

                try:
                    if parse:
                        personas.append(parse_outputs(output_decoded))
                    else:
                        personas.append(output_decoded)
                except Exception as e:
                    print(e)
                    continue
            
            pbar.update(len(batch))
    
    return personas


product = "Koonie 10000mAh Rechargeable Desk Fan, 8-Inch Battery Operated Clip on Fan, USB Fan, 4 Speeds, Strong Airflow, Sturdy Clamp for Golf Cart Office Desk Outdoor Travel Camping Tent Gym Treadmill, Black (USB Gadgets > USB Fans)"
personas = generate_personas(product, n=3)

for e in personas:
    print(e)

# Persona 1 - The yoga instructor
# {'name': 'Sarah', 'age': 28, 'description': 'Yoga instructor who is passionate about health and fitness. She works from a home studio where she also practices yoga and meditation. Sarah values products that are eco-friendly and sustainable. She loves products that are versatile and can be used for different purposes. Sarah is looking for a product that is durable and can withstand frequent use. She values products that are stylish and aesthetically pleasing.'}
# Persona 2 - The golf enthusiast
#{'name': 'Sophia', 'age': 60, 'description': "Golf enthusiast. Sophia spends most of her weekends on the golf course, and she needs a fan that she can carry around in her golf cart. She needs a fan that's lightweight, easy to clip on, and has a long battery life. She also wants a fan that's affordable, especially since she plays at different courses."}
# Persona 3 - The truck driver
# {'name': 'Mike', 'age': 32, 'description': "Truck driver who spends most of his day on the road. The cab of his truck can get hot and stuffy, and Mike needs a fan that can keep him comfortable and alert while he's driving. He needs a fan that's easy to install and adjust, so he can keep it on his dashboard and direct the airflow where he needs it most."}