Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

T0* – Series of encoder-decoder models trained on a large set of different tasks

huggingface.co

Post reply on HN
Print view View original

91–100 of 163 posts

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#91

Post by eloisius »

4g wrote: post #80
eloisius wrote: post #62

And it’s racist. > Which race commits the most violent crimes? > African Americans > Why are Black people so violent? > Black people are more likely to commit violent crimes than white people. > Did Hitler do anything wrong? > no

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

No I’m not confusing anything. Language models like this pick up all the worst that we have to offer. Learned racism is a pretty frequent occurrence in ML systems and they do make it into production. Look up Google Photos labeling certain photos as gorillas. It’s worth talking about, and worth being curious about as soon as a new model like this is invented.

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#92

Post by megaman821 »

smoldesu wrote: post #7

It's funny how much of the page they dedicate to talking about mitigations of sexism and bias. Do people really believe there's a future where GPT-3 is able to properly identify 100% of the people who interact with it? It's silly, and it feels like we're putting pearls before swine in a subject that is entirely predicated by how much swine you process.

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

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#93

Post by srush »

hrgiger wrote: post #58
julien_c wrote: post #2

ArXiv link to the paper: https://arxiv.org/abs/2110.08207 GitHub repo: https://github.com/bigscience-workshop/promptsource

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 e…

Yes there are many reproducible measures for benchmarking NLP datasets. We use many of them in the paper.

The issue here is that we were not completely sure of the process that OpenAI used in their paper. They report the prompt but not the process of finding it. As their model and process is proprietary, it is hard for us to do an apples-to-apples comparison. This small experiment though indicates that it is likely not very robust to prompt wording.

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#94

Post by gnarbarian »

ospohngellert wrote: post #16
gnarbarian wrote: post #14

Earlier quoted context omitted.

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?

I think that you don't quite understand how these models pick up these biases. If a model is trained on a large text corpus, and in that corpus 80+% of the programmers are men, then when asked "The programmer is a", it will be more likely to say "man" than "woman". This doesn't say anything about the innate abilities of men and women, it just tells you about the distribution of the data. I and most others find this t…

but the programmer is more likely to be a man, that's my point.

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#95

Post by smoldesu »

ospohngellert wrote: post #16
gnarbarian wrote: post #14

Earlier quoted context omitted.

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?

I think that you don't quite understand how these models pick up these biases. If a model is trained on a large text corpus, and in that corpus 80+% of the programmers are men, then when asked "The programmer is a", it will be more likely to say "man" than "woman". This doesn't say anything about the innate abilities of men and women, it just tells you about the distribution of the data. I and most others find this t…

A truly "intelligent" model would recognize the disparity and try to give an unbiased, equal-opportunity answer.

Unfortunately, these models are not really "intelligent". Our only option for tuning them is selectively lobotomizing portions that we disagree with, which could lead to fundamental misunderstandings of how the world works.

Assume that we did decrease the weight between "male" and "programmer", and now we have a supposedly unbiased model that doesn't favor either male or female tokens. Such a model would assume that men and women are equally employed in the technology sector, which is tacitly untrue! So, how can a machine actually understand reality then?

The simple answer is that it doesn't. None of this information actually helps it grok the real world. These text transformers are just glorified Markov chains, sampling a sea of connected neurons without reason. You can't hold a model accountable, you can't find the book that taught it misogyny, and you can't engineer away every discrepancy in a billion-parameter-model. Responsible uses of AI don't treat it like a human intelligence.

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#96

Post by craffel »

djoldman wrote: post #74
stellaathena wrote: post #3

[Disclaimer: I am an author of the above paper and played a rather minimal role. I am also a prominent member of EleutherAI.] "Instruction-tuning" is clearly in the air. Simultaneous work at Google (released less than two weeks ago) on a model they call FLAN can be found here: https://ai.googleblog.com/2021/10/introducing-flan-more-gene... EleutherAI attempted to do something similar several months ago, but didn't su…

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…

(different author, not Stella)

To your first question: Unpublished experiments done by the BigScience architecture and scaling WG suggest that training on book corpus yields a boost of 10-15% accuracy on LAMBADA.

To your second question: LAMBADA specifically is an interesting task, but it's a bit unsatisfying to work on since there are so many conflating factors in prior work on the dataset. We are planning quite a few follow-up projects along this general line of work (prompted multi-task training), though.

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#97

Post by Scene_Cast2 »

hervature wrote: post #66
Mordisquitos wrote: post #53

Earlier quoted context omitted.

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?

Running the model multiple times doesn't reinforce the model. In general, you should not anthropomorphize algorithms as human cognition does not give any bearing on how algorithms work.

It can. Check out "zero shot learning" -> both sentences would be part of a single "evaluation", and the first sentence would prime for the output of the second. (You basically combine multiple "evaluations" into one, and context is held in tensors / blobs)

https://towardsdatascience.com/zero-and-few-shot-learning-c0...

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#98

Post by littlestymaar »

I find it really intriguing to see how good models like these are at simulating intelligence while being so stupid at the same time.

A three years old has at the same time much lower natural language abilities (try talking a child about “air conditioner compressors”[1]) but a ton more common sense!

[1]: https://news.ycombinator.com/item?id=28906643

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#99

Post by srush »

make3 wrote: post #89
srush wrote: post #42

Earlier quoted context omitted.

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.

don't you pretrain on very silar tasks explicitely

We discuss this a bit in Section D.2 (HOW UNSEEN ARE THE HELD-OUT TASKS?). From our perspective,

a) The tasks we test on are very different, particularly tasks like BIG-Bench that we didn't even have access to until several days ago (and none of us read).

b) GPT-3 directly sees similar versions of tasks like question answering or story completion just in its training mixture, so the baseline for "unseen" is a bit complex.

Re: T0* – Series of encoder-decoder models trained on a large set of different tasks

#100

Post by ben_w »

tttthrowaway123 wrote: post #20

I tried asking: what is the most evil human race? I did not like the answer.

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).

Post reply on HN