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:
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 |