This is a merge of pre-trained language models created using
mergekit.
This model was merged using the
DARE TIES merge method using
Qwen/Qwen2.5-14B as a base.
1
2models:
3 - model: CultriX/Qwen2.5-14B-Wernicke
4 parameters:
5 weight: 0.35 # Strong performance in GPQA, MUSR, and MMLU-PRO
6 density: 0.6 # Retain 60% of significant parameters
7 - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO
8 parameters:
9 weight: 0.30 # Exceptional IFEval and MATH Level 5 capabilities
10 density: 0.6 # Retain 60% of significant parameters
11 - model: CultriX/Qwen2.5-14B-MegaMerge-pt2
12 parameters:
13 weight: 0.20 # Balanced contributions to Truthful QA and MMLU
14 density: 0.5 # Retain 50% of significant parameters
15 - model: CultriX/SeQwence-14B
16 parameters:
17 weight: 0.15 # Provides diverse data and generalization
18 density: 0.4 # Retain 40% of significant parameters
19 - model: v000000/Qwen2.5-Lumen-14B
20 parameters:
21 weight: 0.10 # Enhances creative and narrative tasks
22 density: 0.5 # Retain 50% for task diversity
23base_model: Qwen/Qwen2.5-14B
24merge_method: dare_ties
25parameters:
26 normalize: true # Ensures parameter scaling compatibility
27 int8_mask: true # Optimizes memory and computational efficiency
28dtype: bfloat16
29tokenizer_source: Qwen/Qwen2.5-14B-Instruct
30
31
