Excellent Model for Writing
This is my current go-to model for writing. It's writing style is slightly better than Gutenberg v3, my previous favorite, and it fixes many of the issues it had. Can you make a gutenberg finetune based on Lyra v3?
Yep, I'll queue that up tonight. Thanks for the feedback and suggestion.
EDIT: actually, I let my colab subscription lapse and I'm currently waiting on new hardware to tune Nemo. I'll have it done in the next week or two.
No problem! Thanks for all that you already do. Noticed that Sao10K/MN-12B-Lyra-v4 has been released. Would it be beneficial to finetune on the newer one instead?
No problem! Thanks for all that you already do. Noticed that Sao10K/MN-12B-Lyra-v4 has been released. Would it be beneficial to finetune on the newer one instead?
I'd like to jump in for a sec to give my 2 cents. I personally prefer v4 myself if Nbeer here is doing another Lyra. V2 and V3 were kinda iffy for me, but V4 has shown some really great promise. It's able to grasp my characters better and doesn't go off the rails as much. It's so good from initial tests that I honestly don't think merging it with something else would make it any better, especially with it being a chatml model on Nemo. DPO/ORPO would be the only thing I can see enhancing V4.
Gutenberg already did wonders for the original Lyra in making it better and more stable, so I'm interested to see if it succeeds here. Even on the UGI leaderboard Lyra v1 with Gutenberg was already one of the better models for writing, beating out the original by quite a bit in some areas, and v4 already matches it in those areas before even adding extra stuff on top.
Lyra4-Gutenberg is on the way. fingers crossed
@ParasiticRogue In your experience, how does it compare Rocinante v1.1 and Mini Magnum/Magnum v2.5?
@ParasiticRogue In your experience, how does it compare Rocinante v1.1 and Mini Magnum/Magnum v2.5?
I haven't spent much time with Drummer's models or Rocinante itself, so I won't get too much into comparisons there. They write well enough for what they are, but from past experience using his others I didn't have much success with them to be honest.
The Magnums though I think are great, but Lyra v4 might edge them out for me at the moment if it keeps up the pace I saw thus far. It just came out, so I can't say for certain yet since I haven't tested in thoroughly, but the main thing that impressed me with v4 right off the bat was it didn't get a character's details wrong in a few scenes which the other models I used had struggled with. It also seems fairly smart for an RP model so far, which kinda shocked me since I usually have to combine them with another model that has a bit more general knowledge to function the way I want. The only thing I didn't care for, and Lyra v1 seemed to have this problem as well, was it can switch between Novel and RP formats with asterisks when the chat isn't set up like that. But that can usually be ironed out when the chat gets longer and the model has a better grasp of what format is in play. Otherwise it seemed like an upgrade from Magnum during initial testing. Haven't done long context yet, but v1 was reported to work well there, so I'm hoping it's the same with this one.
Though again, take these opinions with a grain of salt, since I've only spent an hour or so using it. My stance might change later when I put more effort and time into the model, but it seemed really good in those initial testing scenes.
Sorry for the delay, I've been having issues with my finetuning notebooks crapping out. Will try again tomorrow and work on a better solution for the long-term.
Sorry for the delay, I've been having issues with my finetuning notebooks crapping out. Will try again tomorrow and work on a better solution for the long-term.
Hey, don't sweat it. Shit happens. Hope your setup gets figured out and working good!