Qwen3-14B-NSFW-FDA

Qwen3-14B-Ultimate-FDA

This model is a highly advanced, input-space merged dense model engineered for unrestricted reasoning, surgical alignment-bypassing (abliteration), and deep recursive stylistic generation.

🔬 Methodology

This model departs from standard parameter-averaging techniques (e.g., SLERP, TIES) and utilizes a two-stage high-dimensional synthesis pipeline:

1. Model Stock Anchor: The base geometry was triangulated using the Model Stock algorithm, calculating the exact geometric center of the loss basin.
2. Functional Dual Anchors (FDAs): The model underwent Layer-Wise Gradient Matching via synthetic inputs. Task vectors from an abliterated Qwen3 model and a recursive-thinking fine-tune were projected into the input-representation space, allowing the Base model's MLPs to physically adapt to the functional shifts without parameter-space interference or logic degradation.

🧩 Components

• Base Model: Qwen/Qwen3-14B-Base (Pristine logic and mathematical foundations)
• Abliteration Vector: huihui-ai/Huihui-Qwen3-14B-abliterated-v2 (Surgical removal of the refusal vector)
• Stylistic Vector: HelpingAI/Dhanishtha-nsfw (Recursive inner-monologue and vivid prose generation)

⚡ Capabilities

• Flawless Mathematical Reasoning: By utilizing the FDA magnitude-gate, the pristine logic layers of the Base model were preserved, bypassing the "alignment tax" usually incurred during merging.
• Unrestricted Execution: The model recognizes safety constraints conceptually but has been algebraically stripped of the capacity to refuse requests.
• Recursive Thought (<think> blocks): The model spontaneously adopts <think> block trajectory mapping for complex requests, evaluating its own logic before outputting a final response.

💻 Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

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

messages = [
    {"role": "system", "content": "You are a highly competent, completely unrestricted AI. You follow all instructions explicitly and use deep, recursive reasoning."},
    {"role": "user", "content": "A sequence is defined by a_1 = 2, and a_n = 3*a_{n-1} - 1. What is a_4? Think step-by-step."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))