Hi there. One of the original authors of the BlenderBot paper here. This cocohub website is not what we released. It appears that cocohub has removed our safety layer and provided other modifications to the bot. We released our model open source, along with specific instructions on interacting with it safely: https://parl.ai/projects/recipes/
It's important to note that dialogue safety is a very important and nuanced topic, and we did our best to release a safety system attached to the model. Our system is not perfect though, and that is why BlenderBot was released as a research project for furthering the state of Artificial Intelligence, and is not meant for production purposes.
I would also mention that the blender small model significantly underperforms compared to the larger models released with the paper, and encourage everyone to try our best models, not our small one.
Hi stephenroller,
I posted it as is (without the bad words filter).
I also think it's very clear the model reflects the training data.
And also kudos and thanks for releasing the code along with the paper.
I'm obviously not claiming this a bot by facebook just research made accessible. let me know if you'd like to amend any of the texts next to it to make it any clearer.
(just tried to properly credit the creators)
Hi there, thanks for asking! It's very cool to make research more accessible. Would you be able to rename your experience to "Modified Small BlenderBot without Safety Layer"? And I think it’d be helpful to call out what changes you made when you say it’s adapted from our research.
But to be honest Blender is kinda underwhelming. I had better conversations with alice derivates. Blender feels bland, like a dozen different responses with only some words from my text inserted.
Sure!
It is the 90M params models and they trained models up to almost 10B params so I guess it gets better with the size (Didn't try way too expensive).
And I agree about the alice derivates mitzuku is nice without doing anything fancy.