metadata

base_model: []
library_name: transformers
tags:
  - mergekit
  - merge
license: apache-2.0

Excalibur-7b

Image generated by Envoid's Model9

Magic-Dolphin-7b was an unexpected surprise. Profoundly satisfied with it as a first attempt. For this follow-up I wanted to target the MMLU benchmark specifically. The challenge this time was placing more weight on Merlinite-7b as an unknown quantity that hasn't been in the spotlight despite its novel LAB tuning method.

Excalibur-7b builds on past success and is the culimation of several learnings:

Measuring KL-divergences for new quantization types brought a deeper understanding of benchmarking and assessing model performance
This signifcantly sped up the testing process by using MMLU as a base, narrowing down over 10 candidate linear merges to 1: merliniteX-blockB1
Reaching the limitations of linear merging necessitated a pivot to reviewing the viability of SLERP, dares-ties, and passthrough methods
Thus a competing candidate merge pool was tested between different merge alogrithms. Once more the list was narrowed from 10 candidates to 1: merliniteX-blockF2
merliniteX-blockF2 (SLERP of Magic-Dolphin-7B and jaskier-7b-dpo in unorthadox proportions) was originally planned for release earlier this week
Instead -blockB1 and -blockF2 were merged and the results were placed head to head in a final round of tests. Ultimately a more conventional execution of SLERP showed the best results

output

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

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

models/merliniteX-blockB1
models/merliniteX-blockF2

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
      - model: models/merliniteX-blockF2
        layer_range: [0, 32]
      - model: models/merliniteX-blockB1
        layer_range: [0, 32]
# or, the equivalent models: syntax:
# models:
#   - model: psmathur/orca_mini_v3_13b
#   - model: garage-bAInd/Platypus2-13B
merge_method: slerp
base_model: models/merliniteX-blockF2
parameters:
  t:
    - filter: self_attn
      value: [1, 0.7, 0.3, 0.5, 0]
    - filter: mlp
      value: [0, 0.3, 0.7, 0.5, 1]
    - value: 0.5 # fallback for rest of tensors
dtype: float16