Llama-MiraiFanfare-3.3-70B

Huggingface is capping private repo capacity unless I pay money for limit increase. All test models are now public.
I don't want to pay money for HF that what more or less amounts to just a meme merge.

Anyway, on more intresting news, this model is a tweaked version of the earlier model I did. (The SAO meme model.)
Was getting frustrated with how overly verbose it is, so... This merge was made.

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

This (as of 16/12/24) is now my prefered model to use. Mainly how creative it is. Q2.5 Felt hard to steer.

Prompting

Contrary to popular belief / whats intended for L3 models, ChatML apparently is the way to go. https://huggingface.co/Blackroot/Mirai-70B-1.0/discussions/1#675cc2bc04b8f445a1e62738

Merge Details

Merge Method

This model was merged using the TIES merge method using EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0 as a base.

Models Merged

The following models were included in the merge:

Blackroot/Mirai-70B-1.0

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: Blackroot/Mirai-70B-1.0
    parameters:
      density: 0.4
      weight: 0.35
  - model: EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0
    parameters:
      density: 0.7
      weight: 1

merge_method: ties
base_model: EVA-UNIT-01/EVA-LLaMA-3.33-70B-v0.0
parameters:
  normalize: true
dtype: bfloat16

Llama-3.3-70B-MiraiFanfare

Hourly Usage

Performance Metrics