This is a merge of pre-trained language models created using
mergekit.
Combining
Tissint with
EVA v0.2 seemed to yield nice results in RP, I've uploaded the model so others can try it as well.
Recommended samplers are 0.1 Min P, 1.02 Repititon penalty, and DRY's range set to 2048.
If you'd like to use XTC, I recommend a threshold of 0.2. Lower thresholds seem to adversely affect the coherency.
This model was merged using the della_linear merge method using EVA-UNIT-01/EVA-Qwen2.5-14B-v0.2 as a base.
1models:
2 - model: Ttimofeyka/Tissint-14B-128k-RP
3 parameters:
4 density: 0.4
5 weight: 0.3
6 - model: EVA-UNIT-01/EVA-Qwen2.5-14B-v0.2
7 parameters:
8 density: 0.6
9 weight: 0.7
10
11merge_method: della_linear
12base_model: EVA-UNIT-01/EVA-Qwen2.5-14B-v0.2
13parameters:
14 epsilon: 0.05
15 lambda: 1
16dtype: bfloat16
17
