After a lot of testing by myself and others the general consensus was that Progenitor 1.1 (della_linear) was better than 1.2 (SCE). I actually got some advice from Steelskull about how I had my top k parameter way too high. So I decided to lower it for this merge. Now because I feel 1.2 was still a decent model with the top k on 1, I halved the top k parameter to 0.50 which is basically the max recommended setting. From early testing, I am not sure...it's not blowing me away.
This is a merge of pre-trained language models created using
mergekit.
This model was merged using the
SCE merge method using
nbeerbower/Llama-3.1-Nemotron-lorablated-70B as a base.
1models:
2 # Pivot model
3 - model: SicariusSicariiStuff/Negative_LLAMA_70B
4 # Target models
5 - model: Sao10K/70B-L3.3-Cirrus-x1
6 - model: Sao10K/L3.1-70B-Hanami-x1
7 - model: TheDrummer/Anubis-70B-v1
8 - model: EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.1
9merge_method: sce
10base_model: nbeerbower/Llama-3.1-Nemotron-lorablated-70B
11parameters:
12 select_topk: 0.5
13dtype: bfloat16
