boreas-10_7b-step1

!NB: THIS MODEL NEEDS CONTINUED (PRE)TRAINING

This is the result of step 1 of the upscaling of Boreas-7B with mergekit. It is trying to reproduce the upscaling described in the SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling paper. This model is the result after step 1 from the figure below:

Step 2 continued training is being done - result will be another model.

Merge Details

Merge Method

This model was merged using the passthrough merge method.

Models Merged

The following models were included in the merge:

boreas-7b-8-16-24-32
boreas-7b-16-32
boreas-7b-0-8-16-24
boreas-7b-0-16

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
    - model: boreas-7b-0-16
      layer_range: [0, 16]
  - sources:
    - model: boreas-7b-0-8-16-24
      layer_range: [0, 8]
  - sources:
    - model: boreas-7b-8-16-24-32
      layer_range: [0, 8]
  - sources:
    - model: boreas-7b-16-32
      layer_range: [0, 16]
merge_method: passthrough
dtype: bfloat16

The four models were created with the following configurations:

slices:
  - sources:
    - model: yhavinga/Boreas-7B
      layer_range: [0, 16]
merge_method: passthrough
dtype: bfloat16
---
slices:
  - sources:
      - model: yhavinga/Boreas-7B
        layer_range: [0, 8]
      - model: yhavinga/Boreas-7B
        layer_range: [16, 24]
merge_method: slerp
base_model: yhavinga/Boreas-7B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5
dtype: bfloat16
---
slices:
  - sources:
      - model: yhavinga/Boreas-7B
        layer_range: [8, 16]
      - model: yhavinga/Boreas-7B
        layer_range: [24, 32]
merge_method: slerp
base_model: yhavinga/Boreas-7B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5
dtype: bfloat16
---
slices:
  - sources:
    - model: yhavinga/Boreas-7B
      layer_range: [16, 32]
merge_method: passthrough
dtype: bfloat16