Gemma-2-Ataraxy-v3i-9B

Another experimental model. This one is in the vein of advanced 2.1, but we replace the simpo model used in the original recipe, with a different simpo model, that was more finetuned with writing in mind, ifable. We also use another writing model, which was trained on gutenberg. We use this one at a higher density because SPPO, on paper is the superior training method, to simpo, and quite frankly, ifable is finicky to work with, and can end up being a little too strong.. or heavy in merges. It's a very strong writer but it introduced quite a bit slop in v2.

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

GGUF

https://huggingface.co/lemon07r/Gemma-2-Ataraxy-v3i-9B-Q8_0-GGUF

Merge Details

Merge Method

This model was merged using the della merge method using unsloth/gemma-2-9b-it as a base.

Models Merged

The following models were included in the merge:

• nbeerbower/Gemma2-Gutenberg-Doppel-9B
• ifable/gemma-2-Ifable-9B
• wzhouad/gemma-2-9b-it-WPO-HB

Configuration

The following YAML configuration was used to produce this model:

base_model: unsloth/gemma-2-9b-it
dtype: bfloat16
merge_method: della
parameters:
  epsilon: 0.1
  int8_mask: 1.0
  lambda: 1.0
  normalize: 1.0
slices:
- sources:
  - layer_range: [0, 42]
    model: unsloth/gemma-2-9b-it
  - layer_range: [0, 42]
    model: wzhouad/gemma-2-9b-it-WPO-HB
    parameters:
      density: 0.55
      weight: 0.6
  - layer_range: [0, 42]
    model: nbeerbower/Gemma2-Gutenberg-Doppel-9B
    parameters:
      density: 0.35
      weight: 0.6
  - layer_range: [0, 42]
    model: ifable/gemma-2-Ifable-9B
    parameters:
      density: 0.25
      weight: 0.4

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	21.29
IFEval (0-Shot)	42.03
BBH (3-Shot)	38.24
MATH Lvl 5 (4-Shot)	0.15
GPQA (0-shot)	10.40
MuSR (0-shot)	1.76
MMLU-PRO (5-shot)	35.18