Gemma Advanced V2.1
This is a merge of the 'smartest' advanced fine-tunes available for Gemma-2-9b-it. It includes WPO, SimPO, and SPPO. The merge was performed via the SOTA 'della' merge method. Merge parameters have been hand-tuned for best results. The Q8_0 quant is highly recommended until better quants come along.
Notes and observations:
- The extreme temperature sensitivity from V1 has been fixed, no longer needs to be run at lower temperatures
- Has a somewhat different writing style than any of the parent models
- Great instruction following
- Tracks plot details well and has good situational understanding
- Seems to have a good understanding of psychology, emotions and creative writing
- More 'sane' than base gemma-it, SPPO, or SimPO - not as prone to 'Cruella De Vil' or 'Evil Sorceress' like SPPO or SimPO, when portraying characters
- Would likely serve as a good base for further merges
- I'm looking for a job, if you're hiring. I'm a skilled Python developer who brings strong devops skills along with an ever-growing knowledge of machine learning pipelines and models. Message me if you want to talk about what I can bring to your team.
- Overall, this feels like a very useful and successful merge.
Quantized GGUFs can be found here:
- My quants, Q8_0 tested - jsgreenawalt/gemma-2-9B-it-advanced-v2.1-GGUF
- iMatrix - mradermacher/gemma-2-9B-it-advanced-v2.1-i1-GGUF
- QuantFactory/gemma-2-9B-it-advanced-v2.1-GGUF
- mradermacher/gemma-2-9B-it-advanced-v2.1-GGUF
Thanks to everyone who was kind enough to provide quants!
I'll link to other quants as they appear.
sample ollama Modelfile
FROM /path/to/file/gemma-2-9B-it-advanced-v2.1-Q8_0.gguf
PARAMETER stop "<start_of_turn>"
PARAMETER stop "<end_of_turn>"
PARAMETER num_ctx 8192
TEMPLATE """<start_of_turn>user
{{ if .System }}{{ .System }} {{ end }}{{ .Prompt }}<end_of_turn>
<start_of_turn>model
{{ .Response }}<end_of_turn>"""
This is a merge of pre-trained language models created using mergekit.
Merge Details
Merge Method
This model was merged using the della merge method using google/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:
models:
- model: google/gemma-2-9b-it
- model: wzhouad/gemma-2-9b-it-WPO-HB
parameters:
density: 0.55
weight: 0.6
- model: princeton-nlp/gemma-2-9b-it-SimPO
parameters:
density: 0.35
weight: 0.6
- model: UCLA-AGI/Gemma-2-9B-It-SPPO-Iter3
parameters:
density: 0.25
weight: 0.4
merge_method: della
base_model: google/gemma-2-9b-it
parameters:
normalize: true
int8_mask: true
lambda: 1.0
epsilon: 0.1
dtype: float16