library_name: transformers
tags:
- mergekit
- merge
base_model:
- ifable/gemma-2-Ifable-9B
- jsgreenawalt/gemma-2-9B-it-advanced-v2.1
model-index:
- name: Gemma-2-Ataraxy-v2-9B
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: IFEval (0-Shot)
type: HuggingFaceH4/ifeval
args:
num_few_shot: 0
metrics:
- type: inst_level_strict_acc and prompt_level_strict_acc
value: 21.36
name: strict accuracy
source:
url: >-
https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=lemon07r/Gemma-2-Ataraxy-v2-9B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: BBH (3-Shot)
type: BBH
args:
num_few_shot: 3
metrics:
- type: acc_norm
value: 39.8
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=lemon07r/Gemma-2-Ataraxy-v2-9B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MATH Lvl 5 (4-Shot)
type: hendrycks/competition_math
args:
num_few_shot: 4
metrics:
- type: exact_match
value: 0.83
name: exact match
source:
url: >-
https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=lemon07r/Gemma-2-Ataraxy-v2-9B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GPQA (0-shot)
type: Idavidrein/gpqa
args:
num_few_shot: 0
metrics:
- type: acc_norm
value: 12.3
name: acc_norm
source:
url: >-
https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=lemon07r/Gemma-2-Ataraxy-v2-9B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MuSR (0-shot)
type: TAUR-Lab/MuSR
args:
num_few_shot: 0
metrics:
- type: acc_norm
value: 4.88
name: acc_norm
source:
url: >-
https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=lemon07r/Gemma-2-Ataraxy-v2-9B
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU-PRO (5-shot)
type: TIGER-Lab/MMLU-Pro
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 35.79
name: accuracy
source:
url: >-
https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=lemon07r/Gemma-2-Ataraxy-v2-9B
name: Open LLM Leaderboard
Gemma 2 Ataraxy v2 9B
Finally, after much testing, a sucessor to the first Gemma 2 Ataraxy 9B. Same kind of recipe, using the same principles, same concept as the last Ataraxy but using better models this time.
About
In this merge, we stuck to using models that used preference optimized training (because, while very expensive to train, these are bar none the best performing Gemma finetunes in all my tests), or trained on the amazing gutenberg dataset just like the last one. You can read why jondurbin/gutenberg-dpo-v0.1 is such a good dataset here: https://huggingface.co/lemon07r/Gemma-2-Ataraxy-9B#why-gutenberg.
This time we use the very good advanced 2.1 merge (a merge using the three best preference optimized models), and a new gutenberg model trained on the dataset in the style of SimPO. Both models alone were already better than the original Ataraxy at writing, and general use, which was a pretty high bar to clear. Merging good models, does not always mean a good resulting model. In fact, when the parent models are really good, usually the child model is not as good. This one however, has surprisingly done quite well in my testing thus far and should be a significant upgrade to the last Ataraxy.
GGUF / EXL2 Quants
Bartowski quants (imatrix): https://huggingface.co/bartowski/Gemma-2-Ataraxy-v2-9B-GGUF
Mradermacher quants (static): https://huggingface.co/mradermacher/Gemma-2-Ataraxy-v2-9B-GGUF
Mradermacher quants (imatrix): https://huggingface.co/mradermacher/Gemma-2-Ataraxy-v2-9B-i1-GGUF
Bartowski and mradermacher use different calibration data for their imatrix quants I believe, and the static quant of course uses none. Pick your poison.
More coming soon.
Format
Use Gemma 2 format.
Benchmarks and Leaderboard Rankings
Coming soon.
Merge Details
Merge Method
This model was merged using the SLERP merge method.
Models Merged
This is a merge of pre-trained language models created using mergekit.
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
base_model: ifable/gemma-2-Ifable-9B
dtype: bfloat16
merge_method: slerp
parameters:
t:
- filter: self_attn
value: [0.0, 0.5, 0.3, 0.7, 1.0]
- filter: mlp
value: [1.0, 0.5, 0.7, 0.3, 0.0]
- value: 0.5
slices:
- sources:
- layer_range: [0, 42]
model: jsgreenawalt/gemma-2-9B-it-advanced-v2.1
- layer_range: [0, 42]
model: ifable/gemma-2-Ifable-9B
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 19.16 |
| IFEval (0-Shot) | 21.36 |
| BBH (3-Shot) | 39.80 |
| MATH Lvl 5 (4-Shot) | 0.83 |
| GPQA (0-shot) | 12.30 |
| MuSR (0-shot) | 4.88 |
| MMLU-PRO (5-shot) | 35.79 |