Views
No views yet
1merge_method: dare_ties # Changed to dare_ties
2base_model: CultriX/Qwen2.5-14B-Wernickev3
3dtype: bfloat16 # Use float32 for maximum precision.
4out_dtype: bfloat16 # Output model also uses bfloat16 for consistency and reduced memory usage.
5
6parameters:
7 t: 0.5 # Balances interpolation between models; 0.5 gives equal weight to all contributors.
8 normalize: true # Ensures parameters are normalized to maintain stability during merging.
9 rescale: true # Aligns parameter scales across models for better integration.
10 int8_mask: false # Disable int8 masking to preserve full precision during merging.
11 epsilon: 0.008 # Ultra-fine parameter scaling for precise adjustments between models.
12 lambda: 1.8 # Emphasizes high-impact parameters, giving more weight to significant contributors.
13
14adaptive_merge_parameters:
15 task_weights: # Assign weights to tasks based on their priority and impact on benchmarks.
16 tinyArc: 1.6 # Logical reasoning benchmark; slightly lower priority.
17 tinyHellaswag: 1.5 # Contextual reasoning benchmark with moderate priority.
18 tinyMMLU: 1.8 # Multi-domain knowledge benchmark; important for multitask performance.
19 tinyTruthfulQA: 1.9 # Focuses on factual reasoning and QA; high priority.
20 tinyTruthfulQA_mc1: 1.75 # Multiple-choice factual reasoning; closely related to TruthfulQA.
21 tinyWinogrande: 1.75 # Core reasoning benchmark; slightly lower than BBH.
22 IFEval: 2.30 # Instruction-following tasks; given a high priority for practical applications.
23 BBH: 2.05 # Complex reasoning benchmark; critical for logical tasks.
24 MATH: 2.70 # Highest priority to emphasize mathematical reasoning excellence.
25 GPQA: 2.20 # Graduate-level QA tasks; balanced priority for high-level reasoning.
26 MUSR: 2.15 # Multi-step reasoning; slightly increased to strengthen reasoning performance.
27 MMLU-PRO: 2.00 # Domain multitask benchmark; maintained for general multitask capability.
28 smoothing_factor: 0.03 # Low smoothing for precise task-specific blending without over-generalizing.
29
30gradient_clipping: # Control gradient clipping for each model to stabilize training.
31 CultriX/Qwen2.5-14B-Wernickev3: 0.89 # Higher value ensures stability for the base model.
32 djuna/Q2.5-Veltha-14B-0.5: 0.92 # Stable setting to enhance reasoning contributions.
33 CultriX/SeQwence-14B-EvolMerge: 0.87 # Moderate value for generalist multitask support.
34 qingy2024/Fusion4-14B-Instruct: 0.93 # High stability to emphasize mathematical tasks.
35 CultriX/Qwen2.5-14B-Emerged: 0.88 # Stable setting to maintain multitask performance.
36 sometimesanotion/Lamarck-14B-v0.6: 0.89 # Stable contribution for multi-step reasoning.
37 allknowingroger/QwenSlerp6-14B: 0.90 # Adjusted for stable integration of the replacement model.
38 hotmailuser/QwenSlerp2-14B: 0.91 # Increased slightly for stable integration of reasoning contributions.
39
40models: # Define models to include in the merge, along with their weights and densities.
41 - model: CultriX/Qwen2.5-14B-Wernickev3
42 parameters:
43 weight: 0.33 # Increased to absorb some of the weight from the removed model.
44 density: 0.78 # Maintained optimal density for robust generalist performance.
45
46 - model: djuna/Q2.5-Veltha-14B-0.5
47 parameters:
48 weight: 0.28 # Increased slightly to enhance reasoning benchmarks like MUSR.
49 density: 0.77 # Maintained for strong nuanced reasoning.
50
51 - model: allknowingroger/QwenSlerp6-14B # Replacement for Qwenfinity-2.5-14B.
52 parameters:
53 weight: 0.15 # Matches the weight of the replaced model to preserve balance.
54 density: 0.70 # Increased slightly for stronger parameter integration.
55
56 - model: CultriX/SeQwence-14B-EvolMerge
57 parameters:
58 weight: 0.12 # Moderate weight for general multitask support.
59 density: 0.62 # Maintained for stable contribution.
60
61 - model: qingy2024/Fusion4-14B-Instruct
62 parameters:
63 weight: 0.09 # Moderate weight; focuses on mathematical reasoning tasks.
64 density: 0.75 # Maintained density for stable integration.
65
66 - model: CultriX/Qwen2.5-14B-Emerged
67 parameters:
68 weight: 0.08 # Balanced weight for multitask contributions.
69 density: 0.69 # Maintained density for stable integration.
70
71 - model: sometimesanotion/Lamarck-14B-v0.6
72 parameters:
73 weight: 0.06 # Lower weight to allow more impactful models to dominate.
74 density: 0.62 # Maintained for stable multi-step reasoning contribution.
75
76 - model: hotmailuser/QwenSlerp2-14B
77 parameters:
78 weight: 0.11 # Increased slightly to balance contributions.
79 density: 0.66 # Maintained for stable parameter integration.
80