Views
No views yet
1merge_method: della_linear
2base_model: CultriX/Qwen2.5-14B-Wernickev3
3dtype: bfloat16
4parameters:
5 epsilon: 0.012 # Further reduced to ensure ultra-fine parameter scaling for precision.
6 lambda: 1.4 # Stronger emphasis on significant model contributions.
7 normalize: true # Balances the parameter integration for stability.
8
9adaptive_merge_parameters:
10 task_weights:
11 tinyArc: 1.6 # Prioritizes logical reasoning improvements.
12 tinyHellaswag: 1.5 # Strengthened contextual understanding and consistency.
13 tinyMMLU: 1.65 # Enhanced domain knowledge for multitask benchmarks.
14 tinyTruthfulQA: 1.9 # Maximized for accurate factual reasoning and QA.
15 tinyTruthfulQA_mc1: 1.7 # Balanced focus for multiple-choice reasoning.
16 tinyWinogrande: 1.75 # Advanced reasoning and contextual prediction improvement.
17 IFEval: 1.9 # Instruction-following tasks boosted by multitask contributors.
18 BBH: 1.7 # Complex reasoning is supported by logical base models.
19 MATH: 2.1 # Highest priority, focusing on mathematical excellence.
20 GPQA: 1.8 # Boosted graduate-level QA capabilities.
21 MUSR: 1.9 # Nuanced multi-step reasoning strengthened further.
22 MMLU-PRO: 1.8 # Domain multitask performance maximized.
23 smoothing_factor: 0.1 # Precisely tuned for smooth task-specific blending.
24
25gradient_clipping:
26 CultriX/Qwen2.5-14B-Wernickev3: 0.86 # Backbone stability with slightly reduced clipping.
27 CultriX/Qwenfinity-2.5-14B: 0.83 # Consistent multitask integration.
28 djuna/Q2.5-Veltha-14B-0.5: 0.91 # Strengthened advanced reasoning contributions.
29 CultriX/Qwen2.5-14B-Broca: 0.85 # Logical reasoning enhancements stabilized.
30 qingy2019/Qwen2.5-Math-14B-Instruct: 0.93 # Mathematically focused tasks maximized.
31 CultriX/SeQwence-14Bv1: 0.88 # Generalist multitask support.
32 sometimesanotion/Qwen2.5-14B-Vimarckoso: 0.89 # Balanced multi-step reasoning contributions.
33 allknowingroger/QwenSlerp6-14B: 0.87 # Contextual and logical reasoning integration refined.
34
35models:
36 - model: CultriX/Qwen2.5-14B-Wernickev3
37 parameters:
38 weight: 0.26 # Core backbone for multitask reasoning.
39 density: 0.7 # Slight increase to preserve critical reasoning parameters.
40 - model: CultriX/Qwenfinity-2.5-14B
41 parameters:
42 weight: 0.23 # Comprehensive multitask performer.
43 density: 0.65
44 - model: djuna/Q2.5-Veltha-14B-0.5
45 parameters:
46 weight: 0.22 # Advanced reasoning support for GPQA and MUSR.
47 density: 0.72
48 - model: CultriX/Qwen2.5-14B-Broca
49 parameters:
50 weight: 0.15 # Logical reasoning and factual QA enhancements.
51 density: 0.65
52 - model: qingy2019/Qwen2.5-Math-14B-Instruct
53 parameters:
54 weight: 0.18 # Mathematical reasoning priority.
55 density: 0.73
56 - model: CultriX/SeQwence-14Bv1
57 parameters:
58 weight: 0.14 # Generalist multitask backbone.
59 density: 0.63
60 - model: sometimesanotion/Qwen2.5-14B-Vimarckoso
61 parameters:
62 weight: 0.12 # Multi-step reasoning tasks contributor.
63 density: 0.6
64 - model: allknowingroger/QwenSlerp6-14B
65 parameters:
66 weight: 0.1 # Contextual reasoning improvements.
67 density: 0.62
68
69tokenizer_source: CultriX/Qwen2.5-14B-Wernickev3
70