👁 Impish_Mind_8B

---

---

---

"As she lay her head on the snow-white pillow,
no rest was graced upon her thoughts.

Her impish mind wandered about,
like the mischievous insight that it was.

Her in-dream words, laced with such an impish cleverness
that even a trickster god might have to pause.

Every impish flicker of her reverence—
needed no reason, nor a cause.

Woven with the silken thread of lascivious deeds indeed.

There was no mistaking of her bright—
it was an impish mind all right..."

---

Available quantizations:

• Original: FP16
• GGUF: Static Quants | iMatrix_GGUF | High-Attention | iMatrix-High-Attention
• EXL2: 3.5 bpw | 4.0 bpw | 5.0 bpw | 6.0 bpw | 7.0 bpw | 8.0 bpw
• Specialized: FP8
• Mobile (ARM): Q4_0 | Q4_0_High-Attention

---

TL;DR

• Intended use: Role-Play, Creative Writing, General Tasks.

• Censorship level: Extremely Low

• 9 / 10 (10 completely uncensored)

• UPDATE As of 16.01.2025, Impish_Mind is the highest rated 8B model in the world in the upgraded UGI Leaderboard

• The old score was 8 / 10 and is available for view under images

• Trained on the excellent ruliad/deepthought-8b-llama-v0.01-alpha as base

UGI score:

👁 UGI Score

---

This model was trained with new data and a new approach (compared to my other models). While it may be a bit more censored, it is expected to be significantly smarter. The data used is quite unique, and is also featuring long and complex markdown datasets.

Regarding censorship: Whether uncensoring or enforcing strict censorship, the model tends to lose some of its intelligence. The use of toxic data was kept to a minimum with this model.

Consequently, the model is likely to refuse some requests, this is easly avoidable with a basic system prompt, or assistant impersonation ("Sure thing!..."). Unlike many RP models, this one is designed to excel at general assistant tasks as well.

Regarding low IFeval: I can explain it in depth over 12 pages, but I won't. Instead I'll parse LLAMA 3.3 70B Math eval score:

👁 3.3_Math

---

New abilities (Click on each to expand):

The fun stuff:

• Edgy and paranoid persona, when not in strict assistant mode
• 4chan schizo energy, just the right amount
• A somewhat unique Role Play flavor
• slightly lower positivity bias

---

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

---

Recommended settings for RP

---

Other recommended generation Presets:

---

Your support = more models

My Ko-fi page (Click here)

---

Citation Information

@llm{Impish_Mind_8B,
 author = {SicariusSicariiStuff},
 title = {Impish_Mind_8B},
 year = {2024},
 publisher = {Hugging Face},
 url = {https://huggingface.co/SicariusSicariiStuff/Impish_Mind_8B}
}

---

Benchmarks

---

Other stuff

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

Base model

ruliad/deepthought-8b-llama-v0.01-alpha

(2)

30 items • Updated May 3 • 33