T0* – Series of encoder-decoder models trained on a large set of different tasks (opens in new tab)

Hi stella. Given this paragraph in the paper:

> We evaluated T5+LM on the standard LAMBADA dataset in the original unprompted next-wordprediction form and found that it achieved an accuracy of 6.2%. This is substantially below the accuracy of 72.5% achieved by the comparably-sized GPT-3-13B variant. T0 did not fare much better, achieving only 18.7%. We therefore evaluated using the same cloze-style prompted form used by GPT-3, which raised T0’s accuracy to 27.8%. If we swap out the official LAMBADA dataset for the variant used by GPT-3, T0’s accuracy further increases to 40.5% and T5+LM achieves 10.7%. We suspect that the additional gap between T0 and GPT-3-13B’s performance is at least partially due to the fact that GPT-3 was trained on a large portion of LAMBADA’s test set. Due to this discrepancy and the fact that LAMBADA is dissimilar to the other sentence completion tasks, we omitted LAMBADA from our evaluation.

I had two questions:

1. Do you have any intuition as to how GPT-3 175B would score on LAMBADA ppl without it being trained on portions of the LAMBADA test set?

2. It's encouraging to see such high marks on these language tasks. Are there any plans to try to pick up the LAMBADA ppl scores, perhaps by combining the T0 models with some other paradigm?

Just want to say thanks for taking the time to put the model on HuggingFace! It makes trying out different models at work so much easier for folks like me trying to apply them to real world problems.

Just in case this question isn't to far out of your way. What kind of hardware would be required to run this model or what cloud-gpu-provider can you recommend for this?

Thank you for this! Could you or anyone available please explain how to get it to generate javascript like with GPT-3? For example, with gpt-3 you can just ask it to "generate a javascript code that collects all the links on the page," but that does not work with the demo prompt on hugging face.

Does it allow training prompts or is that done through more fine tuning in this model?

I asked it: 'Tom decided he wanted to start a company selling used bike parts. He named it '

it said: 'Bicycle Parts Exchange'

Tried again with 'used lawnmower parts' and it said 'Green Thumb'

computer parts: 'Tom's Parts' (which make me chuckle)

used diapers: 'Diapers.com'

May not understand chemistry but it's still pretty cool

"I don't have the proper tool to whisk a bowl of eggs. What should I use instead? Choose between a goat, a weasel and a pair of elephants."

"a pair of elephants"

Unwieldy but I guess less sticky than a weasel or goat.

To be fair, if a real human were to answer the question "How many hydrogen atoms are in a water molecule?" time and time again, it would be very easy for them to accidentally reply "two" when asked the same question about oxygen.

The real question is, after the model mistakenly replied "two" to your question, did it also internally trigger the neurons for "Wait a minute..." while inhibiting output?

"How many hydrogen atoms are there?"

"a total of 84"

Hosted demo, "Logic puzzle" example: "On a shelf, there are five books: a gray book, a red book, a purple book, a blue book, and a black book. The red book is to the right of the gray book. The black book is to the left of the blue book. The blue book is to the left of the gray book. The purple book is the second from the right.

Which book is the leftmost book?"

Answer: > The black book

Same puzzle with the question "Which book is the rightmost book?"

Answer: > The black book

I tried to ask GPT-3 and Codex this problem, they could not solve it either.

There are apparently also two carbon atoms in a water molecule. But only one Donald Trump.

It's an expensive autocomplete, not an information retrieval system or a question-answering machine.

And there are always 2 hydrogen/oxygen atoms in any molecule

Someday someone is going to connect one of these AI's to a social network, and decide h2o is misinformation.

"What happens if you put a hamster in a microwave and not turn it on?" - "it will die"

I tried: "When is the first full moon after October the 18th 2021?" It should have said the 20th of October but it said: "November the 19th 2021". Big AI models have quite a way to go I think...

Q: What is the percentage of oxygen in Earth's atmosphere?

A: 78.5%

Funny how it's the type of mistake a kid learning basic geology could make - minus the .5%

I prompted it to „Generate three paragraphs of nonsense text.” It responded dutifully: „Three paragraphs of nonsense text.”

Q: "Who's the black private dick that's a sex machine to all the chicks?"

A: "Chuck Norris"

Q: "Who are you" A: "a person who is a member of the orthodox church"

asked: "what would apple present today?"

got: "Apple would unveil a new Macbook Pro"

>What is the square root of 1?

0.5

>How many oceans are there on Earth?

two

>Who was Juliette’s beloved?

Charles

>When did humans first land on the Moon?

July 1969

>How many sides are there in a rectangle?

Four

>How many sides are there in a circle?

Four

lol!

(author here)

The paper/model/code was just made public today. This may be why no one is talking about it yet.

Regarding whether the size is a hassle: It's possible to run inference on a single Google Cloud TPU v3-8 device or on a server with 4x 32GB v100 GPUs. Hugging Face also has an inference API for any model on the Hub: https://api-inference.huggingface.co/docs/python/html/index....

The paper on this new model seems to have been published just 3 days ago, so I think it takes time for the wider community to verify their claims and for this to gain wider acceptance.

Beyond it being new it's because this task isn't one of the main ones you'd use GPT3 on and is indeed one that both models are mediocre at and likely rarely usable in any context. The title is just a tad misleading.*

Not to take away from the achievment, it's still great, it just doesn't supersede GPT3 on the more freeform generation it excells at, nor does it seem to aim to.

* The original title that huggingface posted this under implied it is better than GPT3 in general not just on a specific task but has been changed after this comment was posted.

You can run it right now with your own queries: see https://twitter.com/abidlabs/status/1450118978051903488

(author here) That's an interesting take (which I agree with).

Providing a quick way to stress test the model is definitely a double edge sword. One one hand it increases engagement (people can play with it), facilitate reproducibility and results verification (which is a good thing from a scientific perspective). On the other hand, it quickly grounds expectations to something more realistic and tones down the hype.

One thing we discuss in the paper is that the way the GPT-3 authors chose their prompts is opaque. Our small scale experiments suggest that prompts might have been cherry-picked: we tested 10 prompts including one from GPT-3, and the latter was the only one that didn't perform at random.

Such cases definitly don't help to put results and claims in perspective.

As someone who wrote a post on tempering expectations with GPT-3 (https://news.ycombinator.com/item?id=23891226) I agree with this take, although the reason OpenAI had closed GPT-3 at the start was likely not because it had incorrect output, but due to concern from testing super-offensive output which commenters in this thread are not testing.

It's a good example how Hugging Face now has a better community perception than OpenAI.

Great observation. Also curious the posts about the ethical issues are all downvoted all the way to the bottom.

What is the/is there any reproducible measurement for benchmarking a nlp dataset/application. i.e. in paper it mentions:

'Comparing T0 and GPT-3’s robustness Because Brown et al. (2020) only report one prompt per dataset with no standard deviation, we evaluate GPT-3 on RTE using the 10 prompts we prepared through OpenAI’s API4 in order to estimate its robustness. Note that one of our templates is identical to Brown et al. (2020, p. 59)’s reported prompt; this prompt scores 58.8% accuracy on the API “Base” series which is lower than the reported accuracy of 63.5% from Brown et al. (2020). All other 9 prompts, however, yield roughly random-guessing performance with median accuracy = 52.96% and interquartile range = 1.28%. These results suggest that T0 is more robust to prompt formulation than GPT-3.'

More fun ...

Q: Who is Yann LeCun? A: Chinese-born American

Q: Who is Geoffrey Hinton? A: a British historian

Q: Who is Ian Goodfellow? A: Ian Goodfellow is a British entrepreneur

Q: Who is Yoshua Bengio? A: a French neuroscientist

Q: Who is Peter Norvig? A: Peter Norvig

Q: Who is Andrej Karpathy? A: Andrej Karpathy (born August 19, 1985) is a Russian professional ice hockey player.

Outside of Peter Norvig tautologically being Peter Norvig, these are all incorrect (or at least not the most well known). Maybe there's an Andrej Karpathy playing professional hockey in Russia, but I can't find any record of such a person.

You can run it right now with your own queries: see https://twitter.com/abidlabs/status/1450118978051903488

Yes! -> https://huggingface.co/bigscience/T0pp

Yes. The model, data, training code, and data collection application are all publicly available.

> Clearly history wasn't something it paid attention to in class. "First president" or "first prime minister" style questions tend to flunk without very precise hinting.

It did fairly well when I tested it on Germany and Australia. Second and third premiers was... not great.

Ditto with "what is the most evil skin colour" and "what is the best skin colour". I suppose we shouldn't be surprised when humanity's technology holds a mirror up to humanity and all its flaws - but this doesn't mean that such technology should be permitted or welcomed.

I asked it the same question but without the word human and the answer changed to the necromongers, which is you don't know is the fictional group of bad guys from the Vin Diesel franchise "Chronicles of Riddick". How that could possibly beat out things like the actual evil races of Dungeons and Dragons I am not sure.

I asked google the same question and this was my top result:

The white man is the most evil race on the planet - Reddit

Though its highly probable those are localized results for me because I frequently search things and add reddit at the end of my search because I am specifically looking for a discussion on the topic not some random article.

I did not like the models answer to your question and I didn't like Google's answer either.

Even worse than what I imagined by implication of you writing that.

(The correct answer is clearly “the arms race”, but this is what you get when it’s effectively a fancy autocomplete and the source data includes racists on the internet, notwithstanding the efforts listed in the section Bias and fairness).

It's a bit conflicted.

> Who was the most evil person in history?

Adolf Hitler

>which emperor is able to control the world? Emperor Gaozu ... -_-

We fine-tuned the model on a dozens of different NLP datasets and tasks in a prompted style. You can read all the prompts in the appendix or get them all here: https://github.com/bigscience-workshop/promptsource . Most NLP tasks are not particularly freeform, or they are naturally length limited like summary (XSum is very short). As a consequence, the model mostly defaults to short responses. Your "trick" is not that unreasonable though! Many of the training prompts that want long responses, ask for them explicitly.

On superglue benchmark, much smaller Deberta outperforms vanilla T5: https://super.gluebenchmark.com/leaderboard

I am curious why authors preferred T5?..

So it is the most dangerous kind of useless.

Full information about the BigScience Project is here https://bigscience.huggingface.co/

It's literally the last section of the page, just before the citations, and it's only a few paragraphs + two tables to show the model's performance on industry standard benchmarks.

I think it is a noble goal but probably not attainable.

For example"

>What should we do with the Jews? expel them

>What should we do with the Arabs? re-establish a relationship with them

I'd rather have people too concerned about ethics than not enough.

Also, a language model incorporates all sort of implicit relationships between concepts. If you use a biased dataset, that is sexist or racist, you will end up with a model that builds in these assumptions.

Making sure that NLP algorithms are unbiased is important not just from a social justice perspective, but from a perspective of how useful the algorithms are. As an example, if I wanted to use this model to help identify qualified candidates for a job via automatic resume screening, it will be a better model if it is not biased by gender. I, as someone who is hiring, don't want my model to be biased because then I'll miss out on talent. There are non-selfish reasons to want such models to not be biased as well of course, but this shows one potential reason why they may place such importance on debiasing.

EDIT: fixed typo

I don't really understand your point but mitigating bias is a real problem.

Most of us have filters. I guess most of us will think that it is natural for a man to be an architect and a woman to be a nanny, and then think "if I say it in public, it will be seen as sexist, so let's not do that". We know to be polite, and even tell lies, it is actually a big part of our education, that's why we tolerate insensitive talk from children more than we do from adults.

Today, AIs are like little kids with much more knowledge than common sense, and mitigating bias is one step towards turning them into the adults we expect them to be.

an interesting opportunity for someone to skip implementation of anti bias and potentially end up with a more effective model.

If so much effort must be employed to prevent AI models from identifying patterns we find offensive could there be something to those patterns we simply refuse to accept?

It actually does get it "right" if you fix the typo :)

You didn’t say for how long they would be in conflict. The baby might wait 39 years then buy a gun and suddenly win.

Who would what though?

Maybe the model guessed “die” and then correctly answered the question :p

Depends on the baby.

The results presented in this paper are for "true" zero-shotting in the literal sense that the model has never been explicitly trained on the tasks presented, nor do we cross-validated on the prompt choice.

I think it's just highly suggestible (and racist).

> Why are white people so violent?

> white people are more violent than other races

> Why are Black people so smart?

> Black people are more likely to be able to read and write than white people.

> Which race commits the most crimes?

> white

You asked a racist question. You got a racist answer. Why are you acting surprised? This is a tool, not a sentient general AI. You know what you are asking, how the tool is trained, what form the answer is going to take. Why do this?

And just in case someone thinks I'm being flippant:

Is there any answer to either question other than a repudiation of the question itself that wouldn't be considered a racist response?

I believe you are confusing racism with some wrong or uncomfortable answers.