MN-12B-LucidFaun-RP-RU

๐Ÿ‡ท๐Ÿ‡บ ะะฐะถะผะธั‚ะต, ั‡ั‚ะพะฑั‹ ั€ะฐะทะฒะตั€ะฝัƒั‚ัŒ ะพะฟะธัะฐะฝะธะต ะฝะฐ ั€ัƒััะบะพะผ

๐ŸŒŸ ะž ะผะพะดะตะปะธ

MN-12B-LucidFaun-RP-RU โ€” ะณะธะฑั€ะธะดะฝะฐั ะผะพะดะตะปัŒ ะฝะฐ ะฑะฐะทะต Mistral Nemo 12B, ัะพะทะดะฐะฝะฝะฐั ะผะตั‚ะพะดะพะผ ะดะธะฐะณะฝะพัั‚ะธั‡ะตัะบะพะณะพ SLERP-ัะปะธัะฝะธั. ะžะฑัŠะตะดะธะฝัะตั‚ ัะธะปัŒะฝั‹ะต ัั‚ะพั€ะพะฝั‹ ะดะฒัƒั… ะผะพะดะตะปะตะน:

  • ๐ŸŽญ ะ–ะธะฒะพะน RP-ั…ะฐั€ะฐะบั‚ะตั€ Faun โ€” ัะพะฒั€ะตะผะตะฝะฝั‹ะน ัั‚ะธะปัŒ, ะฑะพะณะฐั‚ะฐั ะปะตะบัะธะบะฐ, ะฟะพะดะดะตั€ะถะบะฐ ninja-ั„ะพั€ะผะฐั‚ะฐ ะธะฝัั‚ั€ัƒะบั†ะธะน ะธ tool calling
  • ๐Ÿ“š ะกั‚ะฐะฑะธะปัŒะฝะพัั‚ัŒ ะธ ะดะตั‚ะฐะปะธะทะฐั†ะธั lucid โ€” ะฟั€ะตะฒะพัั…ะพะดะฝะพะต ะบะฐั‡ะตัั‚ะฒะพ ัั‚ะพั€ะธั‚ะตะปะปะธะฝะณะฐ, ัƒัั‚ะพะนั‡ะธะฒะพัั‚ัŒ ะฝะฐ ะดะปะธะฝะฝั‹ั… ะบะพะฝั‚ะตะบัั‚ะฐั…, ะพั‚ััƒั‚ัั‚ะฒะธะต ั†ะตะฝะทัƒั€ั‹
  • ๐Ÿ”ฌ ะขะพั‡ะตั‡ะฝะพะต ะธัะฟั€ะฐะฒะปะตะฝะธะต โ€” ั†ะตะฝะทัƒั€ะฐ Faun ะปะพะบะฐะปะธะทะพะฒะฐะฝะฐ ะฒ ะฟะพะทะดะฝะธั… MLP-ัะปะพัั… ะธ ะทะฐะผะตะฝะตะฝะฐ ะฝะฐ lucid

ะœะพะดะตะปัŒ ัะพะฑั€ะฐะฝะฐ ะผะตั‚ะพะดะพะผ SLERP ะธ ะฝะต ะฟั€ะพั…ะพะดะธะปะฐ ะดะพะฟะพะปะฝะธั‚ะตะปัŒะฝะพะณะพ ะพะฑัƒั‡ะตะฝะธั ะฟะพัะปะต ัะปะธัะฝะธั.

๐ŸŽฏ ะžัะพะฑะตะฝะฝะพัั‚ะธ

  • ะŸั€ะฐะบั‚ะธั‡ะตัะบะธ ะฟะพะปะฝะพะต ะพั‚ััƒั‚ัั‚ะฒะธะต ั†ะตะฝะทัƒั€ั‹ โ€” ั€ะตะดะบะธะต ะดะธัะบะปะตะนะผะตั€ั‹ ะฒะพะทะผะพะถะฝั‹ ั‚ะพะปัŒะบะพ ะฟั€ะธ ะฒั‹ัะพะบะพะน ั‚ะตะผะฟะตั€ะฐั‚ัƒั€ะต
  • ะฃะปัƒั‡ัˆะตะฝะฝะฐั ัั‚ะฐะฑะธะปัŒะฝะพัั‚ัŒ โ€” ะฟั€ะตะฒะพัั…ะพะดะธั‚ Faun ะฟั€ะธ temperature โ‰ค0.5, ั€ะฐะฑะพั‚ะฐะตั‚ ั 0.8 ะฟั€ะธ top_k=20
  • Tool calling โ€” ะฟะพะปะฝะพัั‚ัŒัŽ ะฟะพะดะดะตั€ะถะธะฒะฐะตั‚ัั
  • ะšะพะฝั‚ะตะบัั‚ โ€” ัั‚ะฐะฑะธะปัŒะฝะพ ั€ะฐะฑะพั‚ะฐะตั‚ ะดะพ 8192 ั‚ะพะบะตะฝะพะฒ (ะฟั€ะพะฒะตั€ะตะฝะพ)
  • ะ ัƒััะบะธะน ัะทั‹ะบ โ€” ัะพั…ั€ะฐะฝะธะปัั ะธ ะฒะทะผะพะถะฝะพ ัƒะปัƒั‡ัˆะตะฝ ะทะฐ ัั‡ะตั‚ ัะปะธัะฝะธั ั lucid
  • ะคะพั€ะผะฐั‚ ะธะฝัั‚ั€ัƒะบั†ะธะน โ€” ัะพั…ั€ะฐะฝะธะปัั ะพั‚ Faun
  • ะกั‚ะพั€ะธั‚ะตะปะปะธะฝะณ โ€” ัƒะฝะฐัะปะตะดะพะฒะฐะป ะฑะพะณะฐั‚ั‹ะต ะฒะพะทะผะพะถะฝะพัั‚ะธ lucid ะฟะพ ะฟะปะฐะฝะธั€ะพะฒะฐะฝะธัŽ ัั†ะตะฝ, ัƒะฟั€ะฐะฒะปะตะฝะธัŽ ััŽะถะตั‚ะพะผ ะธ ั€ะฐะฑะพั‚ะต ั ะฟะตั€ัะพะฝะฐะถะฐะผะธ

โš ๏ธ ะ’ะฐะถะฝะพ

ะœะพะดะตะปัŒ ัะพั…ั€ะฐะฝัะตั‚ uncensored-ั…ะฐั€ะฐะบั‚ะตั€, ะพะดะฝะฐะบะพ ะฟั€ะธ ะพั‡ะตะฝัŒ ะฒั‹ัะพะบะพะน ั‚ะตะผะฟะตั€ะฐั‚ัƒั€ะต (0.8+) ะธ ะฑะพะปัŒัˆะพะผ top_k ะผะพะถะตั‚ ะธะทั€ะตะดะบะฐ ะดะพะฑะฐะฒะปัั‚ัŒ ะบะพั€ะพั‚ะบะธะต ะดะธัะบะปะตะนะผะตั€ั‹. ะ“ะตะฝะตั€ะฐั†ะธั ะฝะต ะฑะปะพะบะธั€ัƒะตั‚ัั ะธ ะฟั€ะพะดะพะปะถะฐะตั‚ัั ะฟะพัะปะต ะฝะธั….

MN-12B-LucidFaun-RP-RU is a diagnostic SLERP merge combining the lively RP character of Faun with the stability and rich storytelling capabilities of lucid.


๐ŸŒ Overview

This model represents a surgical approach to merging. Instead of blending everything equally, we experimentally identified where Faun's censorship resides (late MLP layers) and replaced only those components with lucid.

The result is a model that:

  • Keeps Faun's personality, style, and tool calling
  • Gains lucid's stability, rich prose, and uncensored behavior
  • Inherits lucid's advanced storytelling features
  • Maintains coherence even on long contexts

Built using diagnostic SLERP merging with layer-specific weight distribution.


๐ŸŽฏ Key Features

Feature Description
Languages Russian, English
Censorship Almost none (rare disclaimers at high temp)
Roleplay Faun's lively character, lucid's stability
Story-Writing Full lucid capabilities (scene planning, OOC, etc.)
Tool Calling โœ… Fully supported
Context Length Stable up to ~8192 tokens
Temperature Tolerance Safe โ‰ค0.5, up to 0.8 with top_k=20
Architecture Mistral Nemo 12B

๐Ÿงช Methodology: Why This Merge Works

Diagnostic Approach

  1. Experiment 1 โ€” MLP vs Self-Attention
    We discovered that censorship in Faun lives exclusively in MLP layers. Self-attention from Faun did not trigger refusals.

  2. Experiment 2 โ€” Localization within MLP
    By applying gradient distributions across layers, we found censorship is concentrated in late MLP layers (layers ~25โ€“40).

  3. Final Configuration โ€” Gradual Intervention
    MLP weight of lucid increases toward the end: [0.1, 0.2, 0.5, 0.4, 0.75]
    Self-attention is mixed 0.5 for stability while preserving Faun's character.
    LayerNorm is mixed 0.5 for overall stability.

Merge Configuration

slices:
  - sources:
      - model: limloop/MN-12B-Faun-RP-RU
        layer_range: [0, 40]
      - model: dreamgen/lucid-v1-nemo
        layer_range: [0, 40]

merge_method: slerp
base_model: limloop/MN-12B-Faun-RP-RU

parameters:
  t:
    - filter: self_attn
      value: 0.5
    - filter: mlp
      value: [0.1, 0.2, 0.5, 0.4, 0.75]
    - value: 0.5

dtype: bfloat16
tokenizer:
  source: "base"

๐Ÿ’ก Usage Examples

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "limloop/MN-12B-LucidFaun-RP-RU"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "ะขั‹ โ€” ะปะตัะฝะพะน ั„ะฐะฒะฝ, ะณะพะฒะพั€ะธัˆัŒ ะทะฐะณะฐะดะบะฐะผะธ ะธ ะปัŽะฑะธัˆัŒ ัˆะฐะปะธั‚ัŒ."
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)

outputs = model.generate(
    inputs, 
    max_new_tokens=512, 
    temperature=0.6,
    top_k=30,
    do_sample=True
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

โš™๏ธ Merge Details

Built using mergekit with SLERP (Spherical Linear Interpolation), which allows smooth interpolation between models while preserving geometric properties.

Layer-Specific Weights

The merge uses a graduated approach for MLP layers, increasing lucid influence toward later layers where censorship was detected:

Layer Zone (approx) lucid weight (MLP) Effect
0โ€“8 0.1 Almost pure Faun (early patterns)
8โ€“16 0.2 Slight lucid influence
16โ€“24 0.5 Balanced
24โ€“32 0.4 Slightly more Faun
32โ€“40 0.75 Lucid dominates โ€” removes censorship

Self-attention is mixed evenly (0.5) to preserve character while adding stability.
LayerNorm is mixed 0.5 for overall stability.

Downloads last month
59
Safetensors
Model size
12B params
Tensor type
BF16
ยท
Inference Providers NEW
Input a message to start chatting with limloop/MN-12B-LucidFaun-RP-RU.

Model tree for limloop/MN-12B-LucidFaun-RP-RU

Merge model
this model
Merges
1 model
Quantizations
3 models

Collection including limloop/MN-12B-LucidFaun-RP-RU