Dear god, this model took half a day to upload because my internet kept cutting out. Totally worth it, though.
This is my attempt at making a more human and verbose version of Aethora v2 by ZeusLabs that's also better at eRP. While I liked both versions of Aethora for its coherency and storytelling for RP, it felt very neutral and tame; not lacking in creativity but restricted in a way. So, why not take that and merge it with two of the most chaotic and human models out there; Sunfall Stheno and Jamet Blackroot. Somehow it worked.
I first created a frankenmerge of Sunfall Stheno and Jamet Blackroot with a bit of Llama 3Some into Blackened Sunfall (check og model card for details). What I'm calling the Blackfall series, merged that with Aethora v2 via the TIES method and you get a bright fluid model while retaining most of the intelligence from individual merged models. Full disclosure, this is an experimental model that needs some more finetuning down the line. Please understand I'm still new to merging and there's probably something I missed. But fow now, I'm thrilled with the results. Proceed with caution.
Quants
GGUF by BackyardAI
GGUF by mradermacher
Merge Details & Recommended Settings
Blackfall Summanus is a good balance between the chaos and literacy. Can get very explicit once in eRP. Still has the Jamet tendency to speak for the user, but is much better at listening to isntructions so but a note in the model instrustions like 'avoid speaking on behalf of {user}' and that'll curb it. Dialogue is much more natural and connotative compared to Aethora v2 but doesn't ramble on like Blackened Sunfall. Understands nuance and context better from being a larger(15B) model and from the additional training in Aethora v2, yet generates more vivid and human text. Likes sticking close to the example text. Leans romantic? The artistic movement, not the genre.
For a merge made from some of the most temperature intolerant models I've seen, Summanus is fine with high temps. Looses coherency past 2, but much better than Blackened Sunfall. Also, this model works best of default template setting. Llama 3 is okay, doesn't have the same 'sparkle' as default but ChatML straight up can break it. I find I get better results with Top P & K, but min-p is great too; lean around 0.08.
Rec. Settings:
Template: Model Default
Temperature: 1.25
Top P: 0.91
Top K: 30
or
Min-P: 0.08
Repeat Penelty: 1.05
Repeat Penelty Tokens: 256
Models Merged
The following models were included in the merge:
Config
The following YAML configuration was used to produce this model:
models:
- model: kromeurus/L3-Blackened-Sunfall-15B
# No parameters necessary for base model
- model: ZeusLabs/L3-Aethora-15B-V2
parameters:
density: [0.25, 0.46, 0.57, 0.38, 0.19]
weight: 0.8
merge_method: ties
base_model: kromeurus/L3-Blackened-Sunfall-15B
parameters:
normalize: true
int8_mask: true
dtype: bfloat16