Update

July, 2025. This model is what started the Impish line of models. It had a lot of "firsts", in several domains. Still holds up very well.

"With that naughty impish grin of hers, so damn sly it could have ensnared the devil himself, and that impish glare in her eyes, sharper than of a succubus fang, she chuckled impishly with such mischief that even the moon might’ve blushed. I needed no witch's hex to divine her nature—she was, without a doubt, a naughty little imp indeed."

Reviews:

TL;DR

• Intended use: Role-Play, General tasks.

• Censorship level: Medium - Low

• 5.5 / 10 (10 completely uncensored)

• UGI score:

"I want some legit RP models of LLAMA 3.2 3B, we got phones!"

"So make one."

"K."

This model was trained on ~25M tokens, in 3 phases, the first and longest phase was an FFT to teach the model new stuff, and to confuse the shit out of it too, so it would be a little bit less inclined to use GPTisms.

It worked pretty well. In fact, the model was so damn thoroughly confused, that the little devil didn't even make any sense at all, but the knowledge was there.

In the next phase, a DEEP QLORA of R = 512 was used on a new dataset, to... unconfuse it. A completely different dataset was used to avoid overfitting.

Finally, another somewhat deep QLORA of R = 128 was used to tie it all together in a coherent way, and connect all the dots, and this was also with a different dataset as well.

The results are sometimes surprisingly good, it even managed to fool some people into thinking it's a MUCH larger model, and sometimes... sometimes it behaves just like you would expect a 3B model to...

Fun fact: the model was uploaded while there were 200 ICBMs headed my way, just flying there in the sky.

I lived, so expect more models in the future!

Available quantizations:

• Original: FP16
• GGUF: Static Quants | iMatrix_GGUF | High-Attention | iMatrix-High-Attention
• EXL2: 4.0 bpw | 5.0 bpw | 6.0 bpw | 7.0 bpw | 8.0 bpw
• GPTQ: 4-Bit-G32
• Specialized: FP8
• Mobile (ARM): Q4_0 | Q4_0_High-Attention

Regarding the format:

It is HIGHLY RECOMMENDED to use the Roleplay \ Adventure format the model was trained on, see the examples below for syntax. It allows for a very fast and easy writing of character cards with minimal amount of tokens. It's a modification of an old-skool CAI style format I call SICAtxt (Simple, Inexpensive Character Attributes plain-text):

SICAtxt for roleplay:

X's Persona: X is a .....
Traits:
Likes:
Dislikes:
Quirks:
Goals:

Dialogue example

SICAtxt for Adventure:

Adventure: <short description>
$World_Setting:
$Scenario:

Model instruction template: Llama-3-Instruct

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{output}<|eot_id|>

Recommended settings for assistant mode

temperature: 0.98

top_p: 0.37

top_k: 100

typical_p: 1

min_p: 0

repetition_penalty: 1.18

do_sample: True

temperature: 1

top_p: 1

top_k: 0

typical_p: 1

min_p: 0.05

repetition_penalty: 1

do_sample: True

temperature: 1.31

top_p: 0.14

top_k: 49

typical_p: 1

min_p: 0

repetition_penalty: 1.17

do_sample: True

temperature: 0.7

top_p: 0.9

top_k: 20

typical_p: 1

min_p: 0

repetition_penalty: 1.15

do_sample: True

Support

• My Ko-fi page ALL donations will go for research resources and compute, every bit is appreciated 🙏🏻

Citation Information

@llm{Impish_LLAMA_3B,
  author = {SicariusSicariiStuff},
  title = {Impish_LLAMA_3B},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/SicariusSicariiStuff/Impish_LLAMA_3B}
}

Benchmarks

Other stuff

• Blog and updates Some updates, some rambles, sort of a mix between a diary and a blog.
• SLOP_Detector Nuke GPTisms, with SLOP detector.
• LLAMA-3_8B_Unaligned The grand project that started it all.