Playing around with GPT-2 - the sequel

Playing around with GPT-2 - the sequel

braintree

In our previous article playing-around-gtp-2 we used the smaller models released by GPT-2.

Now, november 1st 2019, the openai group has launched the full model - and is apparently no longer afraid of the model being abused.

The new larger model is impressively more coherent in its output.

If you have followed the guide in the original article, the new model can be downloaded by running:

python3 download_model.py 1558M

As the model is quite extensive, and the download speed seems throttled - it may well take some time to download the model.

If we run the model on the inputs we tried with the previous smaller model, the difference is quite significant:

python3 src/interactive_conditional_samples.py  --model_name=1558M --top_k 40 --nsamples=2 --length 100

Trying once again with the name of the original article presenting GPT2 we get a impressively more believable feedback: 

Model prompt >>> Language Models are Unsupervised Multitask Learners.
======================================== SAMPLE 1 ========================================


We propose a neural language model (NLLM) for multistep learning that allows models to learn more than one state simultaneously. We focus on the sequential aspect of recurrent neural networks and show that this can provide a powerful approach to learning multistep tasks, such as speech recognition.

In our model, neural processes are represented in the form of a vector, which we feed into the network. Each network layer is a function that takes a vector and returns a matrix containing the
======================================== SAMPLE 2 ========================================


This post is part of a series on Neural Networks with Deep Belief Networks. We describe our architecture and discuss how we use it in our latest project: C2.

When your app starts up, it loads many different modules from different resources. You may need to keep an eye on the models and their dependencies in the app. These models depend on one another, so you may need to watch many of them at once. A typical scenario would look like this:

Model A
================================================================================

For our inquiry on the star wars conundrum the text seems a bit more fluent, and its a lot more coherent although its looks a bit like an excerpt from a star wars fan fic novel:

================================================================================
Model prompt >>> Did Han shoot first ?
======================================== SAMPLE 1 ========================================


Han shot first, so he should be the one who would be punished, right ?

But…

No way… the judge should be the other party!

"You are right, Han shot first." An officer with a serious expression told Feng Nu beside her, his face expressionless as before.

Feng Nu shook her head: "It really didn't matter, Judge, Han is a military officer now. The rules are the same. I do understand
======================================== SAMPLE 2 ========================================


That is a very interesting question, and one I am not sure about myself. To my knowledge, Han shot his first-round shot first. However, at the range, I am certain Han hit another target on at least two different occasions.

I also think his point shot (from his second shot) was the only shot that hit on target.

Why doesn't he have more of these marksmanship feats?

I can only speculate, but here are the possible
================================================================================

On the plot of the new avengers movie we get a bit more details - who the director should be but not a real plot,.

================================================================================
Model prompt >>> What is the plot of the new avengers movie?
======================================== SAMPLE 1 ========================================
 pic.twitter.com/v5pVZW1q5B — Kevin McCorry (@K_McCorry1) March 18, 2016

The new avengers is being directed by Joss Whedon (Dr. Horrible's Sing-Along Blog, Buffy the Vampire Slayer) who co-wrote the upcoming Avengers sequel film with Mark Millar (Kick-Ass, Preacher). It follows the new crop of "Avengers" heroes and villains as they form an unlikely
======================================== SAMPLE 2 ========================================
 We know the new one takes place in Chicago, but what was the city at the first movie?

CRAIG ANTHONY: We had a very open conversation with all our vendors, the way Marvel has their licensing. We asked them to let us know what they would want in a city, and what would their top five choices be, and how they would build a city. We did some extensive research. We went on a map; we went through every street on our own,
================================================================================

GPT2-1558M is a bit unsure on the exact workings of brexit:

================================================================================
Model prompt >>> How about brexit?
======================================== SAMPLE 1 ========================================


We have a Brexit option, that's fine. A no-name "Brexit" option - a bit like the way the euro has been handled by some countries, with some people saying "it wouldn't be as bad as the euro, wouldn't be a catastrophe, would be fine. And then they go on to do it.

The problem with that though, is you want people to vote against their own interests, and if you start to take away options that you want people
======================================== SAMPLE 2 ========================================


The UK's exit from the EU would not immediately lead to a hard border with Irish Republic, but the country's border with Northern Ireland is already a significant bottleneck crossing into either side of the border at Belfast. A hard border would have to be agreed after Brexit.

Image copyright AFP

What happens in a no-deal Brexit?

It is not entirely clear what would happen in a no-deal Brexit. The EU would need to agree a transition deal to avoid a
================================================================================

On the reelection of the Donald in 2020 the model is a sure that he'll get reelected, but is a bit concerned about the attitude towards females in the administration, and seems to think that there is a link to game of thrones.

================================================================================
Model prompt >>> Will donald trump be reelected in 2020 ?
======================================== SAMPLE 1 ========================================
 What role will women play in his administration ? Who will be running his administration when they leave it ? We don't know the answers to many of these questions. And if we go by the conventional wisdom, the answer to none.

A friend of mine, an independent thinker with some interest in politics, just asked me, "Who are the three people he hates the most?" The answer: None. The list includes the President and Vice president.

That's a good question. Why
======================================== SAMPLE 2 ========================================


I think so.

Who's your "dream candidate" ? A political philosopher or a social worker ?

Politicians are like sex. It's really just one way of looking at it.

Follow @SethJames on Twitter.<|endoftext|>Ahead of the release of the long-awaited season 5 of Game of Thrones, it seems we may have a few secrets to add to our list of suspects before we begin guessing who will be the Night King's victims. According
================================================================================

 

Tags