Gemma-2-Ataraxy-9B

Gemma-2-Ataraxy-9B

Working on some new versions. Sit tight.

GGUF / EXL2 Quants

Huge thanks to @mradermacher and @bartowski for making these GGUF quants available to us.

Bartowski quants (imatrix): bartowski/Gemma-2-Ataraxy-9B-GGUF

Mradermacher quants (static): mradermacher/Gemma-2-Ataraxy-9B-GGUF

Mradermacher quants (imatrix): mradermacher/Gemma-2-Ataraxy-9B-i1-GGUF

I think bartowski and mradermacher use different calibration data for imatrix quants, or maybe you prefer static quants. Pick your poison :).

EXL2 quants from CameronRedmore: CameronRedmore/Gemma-2-Ataraxy-9B-exl2

Format

Use Gemma 2 format.

Benchmarks and Leaderboard Rankings

Creative Writing V2 - Score: 82.64 (Rank 1)

https://eqbench.com/creative_writing.html

That's right, much to everyone's surprise (mine included) this model has topped eqbench.com's creative writing benchmark. Big thanks to _sqrkl for taking the time to bother benchmarking this model and adding it ot his leaderboard, and @nbeerbower for entertaining my suggestions for his amazing gutenberg tunes.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

This makes this one of the smartest Gemma 2 9B finetunes/merges, only a few score slightly higher. This model scores above SimPO and SPPO iter2/3 by a whole point on average.

Preface and Rambling

My favorite Gemma 2 9B models are the SPPO iter3 and SimPO finetunes, but I felt the slerp merge between the two (nephilim v3) wasn't as good for some reason. The Gutenberg Gemma 2 finetune by nbeerbower is another my favorites. It's trained on one of my favorite datasets, and actually improves the SPPO model's average openllm leaderboard 2 average score by a bit, on top of improving it's writing capabilities and making the LLM sound less AI-like. However I still liked the original SPPO finetune just a bit more.

Someone suggested that merging the base model on top of the gutenberg may help with tame it back down, which gave me a (possibly) better idea; slerp merging the SimPO finetune on top of the Gutenberg finetune, which is similar to the pretty popular Nephilim v3 recipe, using the Gutenberg finetune in place of the SPPO model, which I thought may give us better results since Gutenberg was trained on top of SPPO.

I wasn't entirely too sure, since if nephilim v3 is anything to go by, it was probably going to also end up worse than the parent models. Normally when I try merges like these, they dont go anywhere. I'm pretty picky, and very skeptical usually, so most times I find that the merge is usually just not better than the original models or only marginally better. Tried this merge anyways to see how it goes, and much to my surprise, this time, I feel like I got very good results. Figured I'd share, and hopefully this wont be just me introducing more useless slop into a world that already has way too many unnecessary merges.

If you're looking for a mistral nemo 12B model instead, I HIGHLY recommend Mistral Nemo Gutenberg v2 by nbeerbower. It's head and shoulders above the many other mistral nemo finetunes I've tried (the first version of mistral nemo gutenburg, romulus simpo, and magnum mini 1.1 being close second favorites).

Why Gutenberg?

We use gutenberg 9b, which is finetuned over SPPO cause of how good the gutenberg dpo dataset is. It's not a german dataset like it sounds, it's a dataset based off of project gutenberg, a public domain collection of popular classic fictions, like Moby Dick, for example. This data set also uses LLM generated responses as the negative prompt, to train models to not sounds so much like AI, or your typical LLMs, and more like actual humans (based on the creative works from project gutenberg). This is quality writing, hence quality data, not just random RP logs, or synthetic data. This dataset, when trained for 3 epochs has shown to increase llama 3 average scores on the old openllm leaderboard, from 72 to 73 (nobody really got higher than 73 before the board revamp), and has already again, proven to increase average openllm 2 leaderboard scores, increasing the average score from 21.47 to 22.61, improving on sppo. Thats a huge improvement. That said, I didnt like gutenberg 9b more than original sppo in real use, felt a tiny bit overfit, so we tried this merge. Did not expect much, because of neph v3 turning out worse than either of its parents, but this surprisingly came out great.

Why is it 10b??

See https://github.com/arcee-ai/mergekit/issues/390

Model is not actually 10b, mergekit is randomly adding lm_head for some reason when doing SLERP merge with Gemma 2 models. I believe Nephilim v3 had a similar issue before they used some sort of workaround that I'm not aware of. Doesn't seem like this affects the GGUF quants, as they're the correct size, so I will leave it as is until mergekit gets a commit that addresses this issue.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• princeton-nlp/gemma-2-9b-it-SimPO
• nbeerbower/gemma2-gutenberg-9B

Configuration

The following YAML configuration was used to produce this model:

base_model: nbeerbower/gemma2-gutenberg-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: princeton-nlp/gemma-2-9b-it-SimPO
  - layer_range: [0, 42]
    model: nbeerbower/gemma2-gutenberg-9B