MedSwin/MedSwin-Merged-DaRE-Linear-KD-0.7

This is a merge of pre-trained language models created using mergekit.

• Developed by: Medical Swinburne University of Technology AI Team
• Funded by: Swinburne University of Technology
• Language(s): English
• License: Apache 2.0

Merge Details

Merge Method

This model was merged using the DaRE-Linear (Drop And REscal in Linear) merging methods, with medalpaca-7b as a base.

• DARE-Linear (or DARE-Task Arithmetic) is the variant where the DARE-processed (sparsified and rescaled) delta parameters are merged using simple linear weighted averaging.
• The final merged model is obtained by adding the weighted sum of the sparsified task vectors back to the base model:
  θ_merged = θ_base + Σ [ α_i * DARE(θ_i - θ_base, p) ], where DARE(τ, p) denotes the operation of dropping parameters with probability p and rescaling the rest by 1/(1-p).

 

Models Merged

The following models were included in the merge:

• MedSwin-7B-KD
• MedSwin-7B-SFT

Configuration

The following YAML configuration was used to produce this model:

base_model: medalpaca-7b
dtype: bfloat16
merge_method: dare_linear
modules:
  default:
    slices:
    - sources:
      - layer_range: [0, 32]
        model: medalpaca-sft
        parameters:
          density: 0.55
          weight: 0.3
      - layer_range: [0, 32]
        model: medalpaca-kd
        parameters:
          density: 0.55
          weight: 0.7
      - layer_range: [0, 32]
        model: medalpaca-7b
parameters:
  rescale: 1.0

Review all model metrics benchmark via Benchmark Document Preview.