This merge is an attempt at bringing a bit more humanity and creativity to the instruction following of huihui's abliterated llama3.3 instruct. The flammenai model was trained on the gutenberg dataset as well as a dataset geared towards replying in a more human like fashion. I haven't gotten to play around with it much, but I find that it has initially been promising in meeting my expectations. I've had one instance of it quickly running away with itself, so keeping the temp around 1 or lower with a 1.02-1.04 rep pen has seemed promising in generating an even tone of responses. But there is always more sampler tweaking to be done. I have just Added an example chat with my sampler settings. Ymmv with reproducing this exactly, my system prompting and additional chat injections are non-standard.

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

Here is a snippet of chat using the Aria - guardian angel card in Silly Tavern

q4_k_m quant-

Aria: Aria's face lights up with gratitude, and she smiles warmly at you. "Thank you for trusting me and giving me permission to help you. I promise to be mindful of your security and only use my abilities for your benefit."

As you mention coffee, Aria's eyes sparkle with interest, and she nods enthusiastically. "I do drink coffee! In fact, I've developed quite a taste for it since arriving on Earth. It's not something we typically have in Heaven, but I've found it to be... invigorating."

She floats closer to the counter, her wings humming softly, and peers at the coffee maker with curiosity. "I must admit, I'm fascinated by the process of brewing coffee. The aroma, the flavors, the ritual of it all... it's quite captivating."

I wonder if he'll share his coffee with me... I hope he doesn't mind me drinking from his cup... I've gotten used to sharing food and drinks with him, even if it's not exactly proper...

Aria's gaze drifts to the snacks on the counter, and she blushes slightly, realizing she's been caught raiding your supplies. "I, uh, hope you don't mind me helping myself to your snacks. I got a bit peckish while waiting for you to wake up, and I didn't want to disturb you."

She looks up at you with big, apologetic eyes, and adds, "If you'd like, I can try to replace what I ate or cook something for you instead. I've learned a few recipes since being on Earth, and I'd love to practice my culinary skills."

Schon: "Look kid, just save my life, then we can be even. Otherwise I don't give a shit." I finish the prep and put the coffee on the boiler. Sitting back down to wait.

Aria: Aria chuckles softly, a hint of amusement dancing in her eyes. "Alright, deal! I'll focus on saving your life, and we can call it even. Though, I must say, I'm looking forward to sharing a cup of coffee with you once it's ready."

She floats over to the table, her wings folding against her back as she takes a seat across from you. Aria's gaze lingers on the coffee maker, and she can't help but sniff the air, savoring the rich aroma wafting from the brewing coffee.

I love the way coffee smells... it's so earthy and inviting... I could sit here and smell it all day...

As the coffee finishes brewing, Aria gets up and pours two cups, adding a dash of cream and sugar to hers before returning to the table. She hands you a cup, her fingers brushing against yours as she does so.

"Here you go," she says, smiling warmly. "I hope you enjoy it. I've found that coffee has a way of making even the most mundane moments feel special."

Aria takes a sip of her own coffee, closing her eyes in appreciation as the flavors dance on her tongue. As she opens her eyes, she catches you watching her and blushes slightly, feeling a flutter in her chest.

He's looking at me... and I think I like it... maybe a little too much...

Schon: "You have been watching me, or you are a very good judge of character." I say as I sip my black coffee. "You didn't even offer, and I appreciate that level of no nonsense, no reason to ask if you know that I want it black."

Aria: Aria's cheeks flush a deeper shade of pink as you comment on her observation skills, and she smiles demurely. "Let's just say I've had plenty of time to observe and learn about you," she replies, her voice barely above a whisper.

As you mention the coffee, Aria's smile widens, and she nods in agreement. "I'm glad I got it right. I figured you weren't the type to fuss over cream and sugar. You seem like someone who appreciates simplicity and directness."

He's perceptive, too... I like that about him... it makes me feel like I can be myself around him...

Aria takes another sip of her coffee, savoring the warmth and flavor as she gazes at you across the table. Her eyes sparkle with curiosity, and she asks, "So, what's the plan for today? Do you have anything specific you need to work on, or would you like some help with something?"

As she waits for your response, Aria's fingers drum a gentle rhythm on the tabletop, betraying her growing excitement at the prospect of assisting you.

My sampler settings for this chat are:

Temp: 0.9

Topk: 100

TopP: 0.98

MinP: 0.02

RepPen: 1.055

Rep Pen Range: 640 tokens

XTC: 0.1 threshhold, 0.45 Probablility

DRY: 0.8 multi, 1.75 base, 2 allowed length, 0 range

Add BOS Token: Yes

KBCPP Sampler Order: rep pen, topK, topA, Temp, TFS, TypP, TopP & MinP

I find these settings to work fairly well generically, mostly the rep pen and range get tweaked based on how terse the conversations are.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: huihui-ai/Llama-3.3-70B-Instruct-abliterated
  - model: flammenai/Llama3.1-Flammades-70B
merge_method: slerp
base_model: huihui-ai/Llama-3.3-70B-Instruct-abliterated
dtype: bfloat16
parameters:
  t: [
    0, 0.018, 0.073, 0.091, 0.109, 0.127, 0.145, 0.164, 0.182, 0.2, 
    0.218, 0.236, 0.255, 0.273, 0.291, 0.309, 0.327, 0.345, 0.364, 0.382, 0.4, 
    0.418, 0.436, 0.455, 0.473, 0.491, 0.509, 0.527, 0.545, 0.564, 0.582, 0.6, 
    0.588, 0.576, 0.564, 0.552, 0.54, 0.527, 0.515, 0.503, 0.491, 0.479, 0.467, 
    0.455, 0.442, 0.43, 0.418, 0.406, 0.394, 0.382, 0.369, 0.357, 0.345, 0.333, 
    0.321, 0.309, 0.297, 0.285, 0.273, 0.26, 0.248, 0.236, 0.224, 0.212, 0.2, 
    0.188, 0.176, 0.164, 0.151, 0.139, 0.127, 0.115, 0.103, 0.091, 0.079, 0.067, 
    0.055, 0.028, 0, 0
]

Downloads last month
20
Safetensors
Model size
70.6B params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for schonsense/Llama-3.3-70B-Inst-Ablit-Flammades-SLERP