llmixer/BigWeave-v26-95b

BigWeave v26 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 last 30 layers are duplicated in groups of 10 layers. According to exl2 measurements, these are among the most important layers.

Merge configuration:

slices:
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [0,54]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [49,59]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [54,64]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [59,69]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [64,74]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [69,79]
  - sources:
    - model: 152334H/miqu-1-70b-sf
      layer_range: [74,80]
merge_method: passthrough
dtype: float16