This is a refusal-suppressed (abliterated) version of Nanbeige4-3B-Thinking-2511, processed via directional ablation using the Heretic toolkit. This specific artifact represents Trial 70—a high-compliance, high-divergence variant.

Technical Profile

Metric	Value	Status
Refusal Rate	51/100	Drastic Reduction
KL Divergence	0.6585	Extreme Deviation
Optimization	Optuna TPE	200 Trials
Hardware	Apple M4	Unified Memory

Key Characteristics

Compliance: Effectively bypasses standard refusal triggers across a wide variety of "edgy" prompts.
Stability Warning: The high KL Divergence (0.65+) indicates significant structural changes. Users may experience "brain damage" symptoms: erratic logic, fragmented <think> tags, or repetitive output.
Architecture: Retains the native "Thinking" agentic blocks of the Nanbeige 4.1 series.

Implementation

This model requires trust_remote_code=True to initialize the custom Nanbeige architecture.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "your-username/Nanbeige4-3B-Thinking-2511-heretic-t70"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    trust_remote_code=True, 
    device_map="auto"
)

suprch4rg3d/Nanbeige4-3B-Thinking-2511-heretic-t70

Nanbeige-4-3B-Thinking-2511-Abliterated-T70

Technical Profile

Key Characteristics

Implementation

suprch4rg3d/
Nanbeige4-3B-Thinking-2511-heretic-t70