File size: 1,758 Bytes
e286bfa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
models:
  - model: CultriX/Qwen2.5-14B-Wernicke
    parameters:
      weight: 0.55         # Backbone model for conversational ability and GPQA
      density: 0.80        # Retain most critical parameters for stability and strength
  - model: VAGOsolutions/SauerkrautLM-v2-14b-DPO
    parameters:
      weight: 0.20         # High IFEval and MMLU-PRO performance with minimized weaknesses
      density: 0.60        # Focus on impactful parameters for specific benchmarks
  - model: rombodawg/Rombos-LLM-V2.6-Qwen-14b
    parameters:
      weight: 0.25         # Enhanced emphasis on reasoning-heavy tasks like MUSR and MATH
      density: 0.70        # Retain reasoning-intensive parameters for improved benchmarks
  - model: allknowingroger/Qwenslerp2-14B
    parameters:
      weight: 0.15         # General stabilizer for consistency across all tasks
      density: 0.65        # Focus on balance and avoiding redundancy
base_model: Qwen/Qwen2.5-14B
merge_method: dare_ties
parameters:
  normalize: true          # Ensure parameter scale consistency
  int8_mask: true          # Optimize for memory and compute efficiency
dtype: bfloat16
tokenizer_source: Qwen/Qwen2.5-14B-Instruct
adaptive_merge_parameters:
  task_weights:
    IFEval: 1.0            # Maintain high IFEval performance
    MATH: 1.3              # Prioritize reasoning and calculation-heavy tasks
    GPQA: 1.1              # Boost factual recall and reasoning accuracy
    MUSR: 1.2              # Enhance logical reasoning and factual understanding
    MMLU-PRO: 1.0          # Retain consistent knowledge representation
  smoothing_factor: 0.15   # Fine-tune blending for stable transitions between tasks
gradient_clipping: 1.0      # Prevent over-contribution from any single model