pipesanma/chasquilla-question-generator

Question Generator

This model should be used to generate questions based on a given string.

Out-of-Scope Use

English language support only.

How to Get Started with the Model

Use the code below to get started with the model.

import torch
from transformers import T5ForConditionalGeneration, T5Tokenizer

def question_parser(question: str) -> str:
    return " ".join(question.split(":")[1].split())

def generate_questions_v2(context: str, answer: str, n_questions: int = 1):
    model = T5ForConditionalGeneration.from_pretrained(
        "pipesanma/chasquilla-question-generator"
    )
    tokenizer = T5Tokenizer.from_pretrained("pipesanma/chasquilla-question-generator")

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = model.to(device)
    text = "context: " + context + " " + "answer: " + answer + " </s>"

    encoding = tokenizer.encode_plus(
        text, max_length=512, padding=True, return_tensors="pt"
    )
    input_ids, attention_mask = encoding["input_ids"].to(device), encoding[
        "attention_mask"
    ].to(device)

    model.eval()
    beam_outputs = model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_length=72,
        early_stopping=True,
        num_beams=5,
        num_return_sequences=n_questions,
    )

    questions = []

    for beam_output in beam_outputs:
        sent = tokenizer.decode(
            beam_output, skip_special_tokens=True, clean_up_tokenization_spaces=True
        )
        print(sent)
        questions.append(question_parser(sent))

    return questions


context = "President Donald Trump said and predicted that some states would reopen this month."
answer = "Donald Trump"

questions = generate_questions_v2(context, answer, 1)
print(questions)

Training Details

Dataset generation

The dataset is "squad" from datasets library.

Check the utils/dataset_gen.py file for the dataset generation.

Training model

Check the utils/t5_train_model.py file for the training process

Model and Tokenizer versions

(v1.0) Model and Tokenizer V1: trained with 1000 rows

(v1.1) Model and Tokenizer V2: trained with 3000 rows

(v1.2) Model and Tokenizer V3: trained with all rows from datasets (78664 rows-train, 9652 rows-validation)