Data Scientist by day, geek tech gamer junky by night.
05/05/2019
Great news from openai regarding their GPT-2 model, they’ve released their 345M parameter version 🙂 GPT-2 Interim Update, May 2019 We’re implementing two mechanisms to responsibly publish GPT-2 and hopefully future releases: staged release and partnership-based sharing. We’re now releasing a larger 345M version of GPT-2 as a next step in staged release, and are sharing the 762M and 1.5B versions with partners in the AI and security communities who…
04/06/2019
Thought I’d ask the Trump Tweets Fine Tuning model a few questions 🙂 All results un-edited Model prompt >>> build a wall to protect against new Mexican people breaking into our country! Wall at our Southern Border is being built by contractors @AgComm Partners & is expected to cost a fortune. So many red tape checks and approvals necessary to build a wall (not an easy task). In January our…
04/06/2019
Currently working on a new bot, gpt-2 117M Fine Tuned on Trump Tweets 🙂 Extracted tweet history from the Trump Twitter Archive, limiting the extract to solely Trump tweets. This gives me over 25k lines to fine tune a model on! It’s currently running and producing some interesting examples, and I can’t honestly tell if they are real or not??? Here are a few examples for now: I know plenty…
04/04/2019
Given what I’ve learned over the past few days, and the success of the Trump bot, I thought I’d try modelling the Mike Hosking data again; extracts below 🙂 This is certainly Hosking’s prose style, although the content matter doesn’t match perfectly. I chose the first question to match what I’d asked the Trump Bot to compare output; happy that it has changed and mirrors Mike build a wall ……
04/04/2019
Today I decided to try and turn the gpt-2 model to Trump’s speeches, and the results are scarily successful. Here is the file of the Trump 2016 Presidential Election Speeches that I used Prompts as titles, with generated Trumpesque speeches 🙂 build a wall Build a wall to stop illegal immigration. I want to work with our allies to put the failed policies of the past behind us. We’ve spent…
03/31/2019
I’ve pivoted with the theme for the text I’m using to fine tune openai’s gpt-2 model, and have picked up a few tips doing so, and I now have Tensorflow gpt-t writing songs 🙂 Steps that I have worked through today: ingesting a collection of lyrics from various artists from a Kaggle soucre encoding and merging the files tweaking the train.py parameters This isn’t going to win any awards. The…
03/30/2019
Using Google’s CoLab [https://colab.research.google.com] I’ve fine tuned the gpt-2 117 model with an extract of text from a kiwi radio host, Mike Hosking I chose Mike not because I ever agree with his opinion, ‘cos he has a distinct writing style. Here’s a link to my GitHub with the files: https://github.com/AmbiguousError/gpt-2 CoLab is symbiotic with Google Drive, and during the fine tune process a large data set that has been…
03/22/2019
We’ve got to start moving a lot more slowly in the next couple years, because the community can be very toxic for anything we do, as well as the actual products at stake. That being said, I think the community is going to need some great leadership development. If you’re working on a project that has some really good data, the people running it will need people to explain what…
03/22/2019
“with this in mind,” said Dr. Eric A. Bowering, an associate professor of physics at UC Davis and a co-author on the paper. “It’s a very good way to learn about particles that cannot be detected in the real world.” Dr. Bowering and four colleagues studied the particles using lasers that were made with molten steel that was sandwiched by an extremely thin thin layer called a layer of carbon.…