flan-t5-large for Extractive QA

This is the flan-t5-large model, fine-tuned using the SQuAD2.0 dataset. It's been trained on question-answer pairs, including unanswerable questions, for the task of Extractive Question Answering.

UPDATE: With transformers version 4.31.0 the use_remote_code=True is no longer necessary.

This model was trained using LoRA available through the PEFT library.

NOTE: The <cls> token must be manually added to the beginning of the question for this model to work properly. It uses the <cls> token to be able to make "no answer" predictions. The t5 tokenizer does not automatically add this special token which is why it is added manually.

Overview

Language model: flan-t5-large
Language: English
Downstream-task: Extractive QA
Training data: SQuAD 2.0
Eval data: SQuAD 2.0
Infrastructure: 1x NVIDIA 3070

Model Usage

Using Transformers

This uses the merged weights (base model weights + LoRA weights) to allow for simple use in Transformers pipelines. It has the same performance as using the weights separately when using the PEFT library.

import torch
from transformers import(
  AutoModelForQuestionAnswering,
  AutoTokenizer,
  pipeline
)
model_name = "sjrhuschlee/flan-t5-large-squad2"

# a) Using pipelines
nlp = pipeline(
  'question-answering',
  model=model_name,
  tokenizer=model_name,
  # trust_remote_code=True, # Do not use if version transformers>=4.31.0
)
qa_input = {
'question': f'{nlp.tokenizer.cls_token}Where do I live?',  # '<cls>Where do I live?'
'context': 'My name is Sarah and I live in London'
}
res = nlp(qa_input)
# {'score': 0.984, 'start': 30, 'end': 37, 'answer': ' London'}

# b) Load model & tokenizer
model = AutoModelForQuestionAnswering.from_pretrained(
  model_name,
  # trust_remote_code=True # Do not use if version transformers>=4.31.0
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

question = f'{tokenizer.cls_token}Where do I live?'  # '<cls>Where do I live?'
context = 'My name is Sarah and I live in London'
encoding = tokenizer(question, context, return_tensors="pt")
output = model(
  encoding["input_ids"],
  attention_mask=encoding["attention_mask"]
)

all_tokens = tokenizer.convert_ids_to_tokens(encoding["input_ids"][0].tolist())
answer_tokens = all_tokens[torch.argmax(output["start_logits"]):torch.argmax(output["end_logits"]) + 1]
answer = tokenizer.decode(tokenizer.convert_tokens_to_ids(answer_tokens))
# 'London'

Metrics

# Squad v2
{
    "eval_HasAns_exact": 85.08771929824562,
    "eval_HasAns_f1": 90.598422845031,
    "eval_HasAns_total": 5928,
    "eval_NoAns_exact": 88.47771236333053,
    "eval_NoAns_f1": 88.47771236333053,
    "eval_NoAns_total": 5945,
    "eval_best_exact": 86.78514276088605,
    "eval_best_exact_thresh": 0.0,
    "eval_best_f1": 89.53654936623764,
    "eval_best_f1_thresh": 0.0,
    "eval_exact": 86.78514276088605,
    "eval_f1": 89.53654936623776,
    "eval_runtime": 1908.3189,
    "eval_samples": 12001,
    "eval_samples_per_second": 6.289,
    "eval_steps_per_second": 0.787,
    "eval_total": 11873
}

# Squad
{
    "eval_HasAns_exact": 85.99810785241249,
    "eval_HasAns_f1": 91.296119057944,
    "eval_HasAns_total": 10570,
    "eval_best_exact": 85.99810785241249,
    "eval_best_exact_thresh": 0.0,
    "eval_best_f1": 91.296119057944,
    "eval_best_f1_thresh": 0.0,
    "eval_exact": 85.99810785241249,
    "eval_f1": 91.296119057944,
    "eval_runtime": 1508.9596,
    "eval_samples": 10657,
    "eval_samples_per_second": 7.062,
    "eval_steps_per_second": 0.883,
    "eval_total": 10570
}

Using with Peft

NOTE: This requires code in the PR https://github.com/huggingface/peft/pull/473 for the PEFT library.

#!pip install peft

from peft import LoraConfig, PeftModelForQuestionAnswering
from transformers import AutoModelForQuestionAnswering, AutoTokenizer
model_name = "sjrhuschlee/flan-t5-large-squad2"