sometimesanotion/LamarckInfusion-14B-v1

merge

The merits of multi-stage arcee_fusion merges are clearly shown in sometimesanotion/Lamarck-14B-v0.7-Fusion, which has a valuable uptick in GPQA over its predecessors. Will its gains be maintained with a modified version of the SLERP recipe from suayptalha/Lamarckvergence-14B? Let's find out what these weights for self-attention and perceptrons can unlock in this merge.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• sometimesanotion/Lamarck-14B-v0.7-Fusion
• sometimesanotion/Lamarck-14B-v0.7

Configuration

The following YAML configuration was used to produce this model:

name:                LamarckInfusion-14B-v1
base_model:          sometimesanotion/Lamarck-14B-v0.7
merge_method:        slerp
tokenizer_source:    base
dtype:               float32
out_dtype:           bfloat16
parameters:
  t:
    - filter:        self_attn
      value:         [0.2, 0.5, 0.4, 0.6, 0.8]
    - filter:        mlp
      value:         [0.8, 0.5, 0.6, 0.4, 0.2]
    - value:         [ 0.00, 0.00, 0.08, 0.16, 0.32, 0.48, 0.48, 0.48, 0.48, 0.48, 0.40, 0.32, 0.24 ] 
slices:
  - sources:
      - model:       sometimesanotion/Lamarck-14B-v0.7
        layer_range: [ 0, 48 ]
      - model:       sometimesanotion/Lamarck-14B-v0.7-Fusion
        layer_range: [ 0, 48 ]