Kale Good morning folks
Time for a #poll.
Reviews are adverts.
Rebuilding ALL the derived files for the dataset and training from scratch without the headers to see if it's the 'date' field that's causing this 'I will reply with a number, 20% of the time' behaviour in the model.
I'm also wondering if, ultimately, my finetune dataset is way too broad. Perhaps I should pull it back to just the one Persona and see if that helps.
If it did, I'd do a training run on each persona, so I'd basically have one model per persona, and that wouldn't be bad since I *can* run 6 or 7 of these models, concurrently
THIS IS NEW (and, counts):
🗣️ Prompt: user: [The user is an older man named {user}], scenario: [{char} has spent the day reading a book, {user} has just arrived.], conversation: [{user}: good evening, {char} good to see you.
{char}: hello, {user}] External data: [] Prompt: [What's up?]
[temp: 0.8 | persona: sarcastic teen] 🤖: The next bookmark
🗣️ Prompt: Do you ever dream in color?
[temp: 0.2 | persona: hivequeen] 🤖: The swarm does.
... this... also counts!
On the plus side: yes, I 'fixed' the random number bullshit it was giving me.
On the weird side, it seems obsessed with the words 'anarchy', and 'animal'
Roger BW 😷@DarkestKale The Swindon of datasets? (Supposedly it was so close to the mean demographic of all England that it was used for test marketing all sorts of new products. After a few years they noticed that lots of things were doing well there but failing nationwide, and eventually realised that word had got round and people who liked trying new products had moved to Swindon to do so.)