This is a merge of pre-trained language models created using
mergekit.
This model was merged using the
Linear DELLA merge method using
Sakalti/ultiima-78B as a base.
1merge_method: della_linear
2base_model: Sakalti/ultiima-78B
3dtype: float16
4parameters:
5 epsilon: 0.015 # Fine-grain scaling for precision.
6 lambda: 1.6 # Strong emphasis on top-performing models.
7 normalize: true # Stable parameter integration across models.
8adaptive_merge_parameters:
9 task_weights:
10 tinyArc: 1.75 # Logical reasoning.
11 tinyHellaswag: 1.65 # Contextual predictions.
12 tinyMMLU: 1.8 # Domain knowledge.
13 tinyTruthfulQA: 2.0 # Prioritize truthful reasoning.
14 tinyTruthfulQA_mc1: 1.85
15 tinyWinogrande: 1.9 # Advanced reasoning and predictions.
16 IFEval: 2.1 # Instruction-following and multitasking.
17 BBH: 1.9 # Complex reasoning.
18 MATH: 2.3 # Mathematical reasoning.
19 GPQA: 2.2 # Factual QA.
20 MUSR: 2.0 # Multi-step reasoning.
21 MMLU-PRO: 2.2 # Domain multitask performance.
22 smoothing_factor: 0.1 # Smooth blending across benchmarks.
23models:
24 - model: MaziyarPanahi/calme-2.4-rys-78B
25 parameters:
26 weight: 1
27 density: 1
28 - model: Sakalti/ultiima-78B
29 parameters:
30 weight: 1
31 density: 1
