Llama_3.x_70b_Trojka_V2

about

A series of experiment of "empowerement" of models with my usual stabilizator (L3.1 70B Hermes 3 lorablated and its finetunes) and recently discovered perplexity-dropper (L3.1 70B Tess 3)

• This version is based on SteelSkull's Electra R1, my new favorite merge.
• And it gives, benchmark-wise, good results, especially on ARC-C

---

benchs

• PPL 512 Wikitext Eng : 3.10 (very good)
• ARC-C : 65.90 (I never merged anything >= 64 before, and that's already very high, most of L3 merges are between 53 and 59. 64 +3% is beyond the margin of error, this might be a record.)
• ARC-E : 83.85 (This is the usual plateau of my best merges)

---

merge

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

Merge Details

Merge Method

This model was merged using the Model Stock merge method using Steelskull/L3.3-Electra-R1-70b as a base.

Models Merged

The following models were included in the merge:

• mlabonne/Hermes-3-Llama-3.1-70B-lorablated
• migtissera/Tess-3-Llama-3.1-70B

Configuration

The following YAML configuration was used to produce this model:

merge_method: model_stock
models:
  - model: mlabonne/Hermes-3-Llama-3.1-70B-lorablated
    parameters:
      weight: 1.0
  - model: migtissera/Tess-3-Llama-3.1-70B
    parameters:
      weight: 1.0
base_model: Steelskull/L3.3-Electra-R1-70b
dtype: bfloat16
out_dtype: bfloat16
parameters:
  int8_mask: true
  normalize: true
  rescale: false
chat_template: auto
tokenizer:
  source: union

Same as

merge_method: model_stock
models:
  - model: Steelskull/L3.3-Electra-R1-70b
    parameters:
      weight: 1.0
  - model: mlabonne/Hermes-3-Llama-3.1-70B-lorablated
    parameters:
      weight: 1.0
  - model: migtissera/Tess-3-Llama-3.1-70B
    parameters:
      weight: 1.0
base_model: Steelskull/L3.3-Electra-R1-70b
dtype: bfloat16
out_dtype: bfloat16
parameters:
  int8_mask: true
  normalize: true
  rescale: false
  filter_wise: false
chat_template: auto
tokenizer:
  source: union

Same as

merge_method: model_stock
models:
  - model: Steelskull/L3.3-Electra-R1-70b
    parameters:
      weight: 1.0
  - model: migtissera/Tess-3-Llama-3.1-70B
    parameters:
      weight: 1.0
  - model: mlabonne/Hermes-3-Llama-3.1-70B-lorablated
    parameters:
      weight: 1.0
base_model: Steelskull/L3.3-Electra-R1-70b
dtype: bfloat16
out_dtype: bfloat16
parameters:
  int8_mask: true
  normalize: true
  rescale: false
  filter_wise: false
chat_template: auto
tokenizer:
  source: union