@blakehunsicker yes, it can be fine-tuned at a rate of ~5000 tokens/second, which should be sufficient for small-to-medium-size datasets. Fine tuning instructions are here: https://github.com/kingoflolz/me...
@pallpakk some results were definitely weird but overall, it works great! Negative sentiment, foul language, etc are context specific outputs. So if an input is negative/abusive itself, the output is bound to reinforce the same sentiment.
Hey hope this is still relevant
I find gpt-j quite alright in the generation but it provides silly results when it does summaries. are there any experts here that could help on how i can maybe train here to provide tl;dr's
Yep! Doors have been OPENED 🤯 An open-source cousin of GPT-3 is here 😇
- Performs on par with 6.7B GPT-3
- Performs better and decodes faster than GPT-Neo
- repo + colab + free web demo
Got to know about it through Towards Data Science article: https://towardsdatascience.com/c...
More details in @arankomatsuzaki's article: https://arankomatsuzaki.wordpres...