llmixer/BigWeave-v25-95b

BigWeave v25 95b

The BigWeave models aim to experimentally identify merge settings for increasing model performance. The version number merely tracks various attempts and is not a quality indicator. Only results demonstrating good performance are retained and shared.

Prompting Format

Chatml, Mistral, Vicuna.

Merge process

This is a self-merge of 152334H/miqu-1-70b-sf. The first 30 layers are duplicated in groups of 10 layers. According to exl2 measurements, these are among the least important layers.

Merge configuration:

slices:
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [0,6]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [1,11]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [6,16]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [11,21]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [16,26]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [21,31]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [26,80]
merge_method: passthrough
dtype: float16