Macaw (Multi-angle c(q)uestion answering) is a ready-to-use model capable of general question answering, showing robustness outside the domains it was trained on. It has been trained in "multi-angle" fashion, which means it can handle a flexible set of input and output "slots" (like question, answer, explanation) .

Macaw was built on top of T5 and comes in different sizes: macaw-11b, macaw-3b, and macaw-large, as well as an answer-focused version featured on various leaderboards: macaw-answer-11b (see below).

Some suggestive examples from the Macaw (11B) model, for different angles:

• (Q→A) Given a question, what's the answer?
  Q: James went camping in the woods, but forgot to bring a hammer to bang the tent pegs in. What else might he use?
  → A: rocks

• (QM→A) Given a question and answer choices, what's the answer?
  Q: James went camping in the woods, but forgot to bring a hammer to bang the tent pegs in. What else might he use?
  M: (A) a leaf (B) a log (C) a worm
  → A: a log

• (Q→AE) Given a question, what's the answer and an explanation?
  Q: Which force pulls objects to the ground?
  → A: gravity
  → E: Gravitational force causes objects that have mass to be pulled down on a planet.

• (A→QE) Given an answer, what's a plausible question and explanation?
  A: elephant
  → Q: Which animal has the largest ears?
  → E: The ears of an elephant are the largest.

• (C→QA) Given a context, what's a plausible question and answer?
  C: A car needs a battery to start.
  → Q: What is required for a car to start?
  → A: battery

For many more examples of the basic Q→A angle, see examples.md.

Macaw can easily be used in the Hugging Face transformers library, as shown here for the smallest model (the smallest model is not generally recommended, but has much smaller footprint), where given a question we want to return an answer and suggested multiple-choice answer options.

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("allenai/macaw-large")
model = AutoModelForSeq2SeqLM.from_pretrained("allenai/macaw-large")
input_string = "$answer$ ; $mcoptions$ ; $question$ = What is the color of a cloudy sky?"
input_ids = tokenizer.encode(input_string, return_tensors="pt")
output = model.generate(input_ids, max_length=200)

>>> tokenizer.batch_decode(output, skip_special_tokens=True)
['$answer$ = gray ; $mcoptions$ = (A) blue (B) white (C) grey (D) white']

(run pip install -r requirements.txt if any dependencies are missing). Note there's no guarantee the different slots are fully coherent, as in gray/grey (and duplicate "white") here, more so for the macaw-large model vs the larger ones.

The code in macaw/utils.py includes some convenience wrappers, such as load_model and run_macaw, here are some examples loading the macaw-11b model onto two GPUs (need around 48GB total GPU memory for the largest model to work):

from macaw.utils import load_model, run_macaw
model_dict = load_model("allenai/macaw-11b", cuda_devices=[0,1])
res1 = run_macaw("Q: Which force pulls objects to the ground?\nA\nE", model_dict)
# Alternate input syntax
res2 = run_macaw({"Q:":"Which force causes a compass needle to point north?", "A":""}, model_dict)
# Add sampling options for the output
res3 = run_macaw("Q: Which force pulls objects to the ground?\nA\nE", model_dict, {"do_sample": True, "temperature": 2.0})

>>> [print(res["output_slots_list"][0]) for res in [res1, res2, res3]]
{'answer': 'gravity', 'explanation': 'Gravitational force causes objects that have mass to be pulled down on a planet.'}
{'answer': 'magnetism'}
{'answer': 'gravitional force', 'explanation': 'Gravitational force causes objects that have mass to be pulled down on a planet.'}

For batch evaluation of instances at various angles, see macaw/batch_eval.py for pointers.

Here are the slots available in Macaw, generally applicable for both input and output:

An angle is a specific set of input/output slots, for instance QM->AE is the task of producing answer and explanation, given a question and multiple-choice options. Macaw is trained on a wide variety of angles and handles unseen angles as well, one exception is that the context (C) only appears as an input slot in the training data.

The Challenge300 dataset of 300 diverse probing examples can be found in challenge300-probes-v1.jsonl. The basic Q→A output from Macaw (at different sizes), as well as outputs from GPT3, Jurassic-1 and alternate T5 models trained on NaturalQuestions, can be seen in examples.md.

See DEMO.md for instructions and code to host an interactive version of Macaw.

Macaw was trained in two steps from the text-to-text transformer model T5:

1. Multi-angle version of UnifiedQA by fine-tuning T5 on the following 7 datasets and associated angles:

   • BoolQ, SQuAD2.0, NarrativeQA: QC→A, AC→Q
   • ARC, OBQA: QMC→A, QC→A, QM→A,QAC→M, MAC→Q, AC→QM
   • RACE, MCTest: QMC→A, QC→A, QAC→M,MAC→Q

2. Further fine-tuning of Multi-Angle UnifiedQA on multiple-choice and direct-answer elementary science questions, along with (up to 5) explanation sentences from WorldTreeV2:

   • ARC: QMC→AE, AQC→M, QMEC→A, QME→A, QE→A, QMC→A, QC→AE, QM→AE, QMAC→E, QMA→E
   • ARC-DA: QC→AE, Q→AE, QC→A, Q→A, QEC→A, QE→A, AE→Q, AC→Q, QA→E, AQC→E

3. A specialized answer-focused model, macaw-answer-11b (called "UnifiedQA + ARC MC/DA + IR" on the leaderboards for ARC, ARC-Easy, and ARC-DA) was trained on a smaller set of angles, not including explanations:

   • ARC: QMC→A, QAC→M, QC→A, QM→A, MAC→Q, AC→QM, M→QA
   • ARC-DA: QC→A, Q→A, AC→Q, C→QA

The Macaw models can be accessed from the Hugging Face model hub:

• macaw-11b (11 billion parameters)
• macaw-3b (3 billion parameters)
• macaw-large (770 million parameters)
• macaw-answer-11b (11 billion parameters)

For a sense of the degradation in performance for the smaller sizes, here are baseline scores on the ARC Challenge and ARC Easy multiple-choice development questions. Included are variants with and without IR context from a large science corpus (corresponding to angles QMC→A and QM→A respectively).

As a model capable of generating free form text, the output of the model is not guaranteed to be free of offensive material, so appropriate caution is advised when using the model.

If you use Macaw in your work, please reference the related paper using

@article{Tafjord2021Macaw,
  title={General-Purpose Question-Answering with {M}acaw},
  author={Oyvind Tafjord and Peter Clark},
  journal={ArXiv},
  year={2021},
  volume={abs/2109.02593}
}