sensix-zo/Gemma-4-31B-Paite-Uncensored-16bit

Sensix Paite 31B Master CPT (16-bit Full Weights)

This repository hosts the full 16-bit merged weights of the Sensix Paite 31B model. This model is a high-intelligence foundation developed by merging a high-rank (r=128) Continued Pre-Training (CPT) adapter into the Gemma-4-31B architecture. It is specifically optimized for native Paite linguistic fluency while maintaining the sophisticated reasoning and unfiltered persona of the original instruction-tuned base.

Architecture and Development

The creation of this model involved a specific architectural modification to the standard Gemma 4 framework to ensure optimal weight distribution and compatibility.

Structural Unwrapping

The base Gemma 4 model utilizes "Clippable" linear layers which can interfere with standard fine-tuning and weight merging. During the CPT phase, all Gemma4ClippableLinear modules were unwrapped and replaced with standard linear layers. This allows for a clean 16-bit merge where the specialized Paite linguistic weights are fused directly into the core matrices of the 31B parameter stack.

Specialized Thinking Framework

This model utilizes the Gemma 4 "Thinking" protocol. It is capable of utilizing a dedicated internal monologue channel to process complex linguistic nuances and logical sequences before generating a final response. This makes it exceptionally capable of translating complex English technical or philosophical concepts into the Paite language without losing logical density.

Model Features

• Native Paite Proficiency: Mapped via the Paite Bible and modern 2025-12-13 vocabulary updates.
• Uncensored Heretic Logic: Inherits an unfiltered, direct-response framework that follows instructions without safety-refusal bottlenecks.
• High-Rank Adaptation: Developed with Rank 128/Alpha 128 LoRA to ensure deep weight penetration during the CPT phase.
• Full 16-bit Precision: Merged in bfloat16 to preserve the maximum intelligence and nuance of the 31B parameter model.

Technical Specifications

• Base Model: Gemma 4 31B Instruction Tuned (Grand Horror X Thinking)
• Total Parameters: 31 Billion
• Precision: bfloat16
• Context Window: 4096 Tokens
• Language Support: Paite (pck) and English (en)
• Inference Style: Thinking / Multi-turn Dialogue

Usage and Inference

This is a full-weight model. It can be loaded using standard Transformers or the Unsloth library for accelerated performance.

Loading via Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "sensix-zo/Gemma-4-31B-Paite-Master-16bit"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

Prompting with Thinking Channel

To activate the internal reasoning process, use the chat template and ensure the generation prompt is set to true.

messages = [
    {"role": "user", "content": "Paite pau in, 'The impact of technology on society' chungtang thulim khat gelh in."}
]

prompt = tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True, 
    enable_thinking=True
)

inputs = tokenizer(text=prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.8)
    print(tokenizer.decode(outputs[0], skip_special_tokens=False))

Training Data

The model underwent Continued Pre-Training using a curated dataset of Paite linguistic structures and high-IQ reasoning chains. The dataset was cleaned and packed to ensure the model learned grammatical precision while maintaining the "Intense" and "Uncensored" nature of the base model.

The CPT phase utilized the Unsloth library to manage the high-rank adapter training before the final 16-bit weight fusion.

Ethics and Disclaimer

This model is UNCENSORED and INTENSE. It is designed to be a tool for linguistic research and advanced reasoning. It does not possess the standard safety guardrails found in mainstream AI models. The developers are not responsible for the content generated. Users are expected to comply with their local laws and use the model at their own risk.