aifeifei798/DarkIdol-Gemma-4-31B-it-FT

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

DarkIdol-Gemma-4-31B-it-FT by aifeifei798 is a 31 billion parameter Gemma-4 based instruction-tuned model with a 32768 token context length. This model is specifically engineered for research into the "Alignment Tax" in LLMs, explicitly revealing internal safety markers (ESMs) rather than suppressing them. It excels in role-playing scenarios, creative writing, and provides faster inference by orthogonalizing refusal vectors, making it suitable for studying the friction between AI intelligence and safety mechanisms.

Loading preview...

DarkIdol-Gemma-4-31B-it-FT: Researching the "Alignment Tax"

This model, developed by aifeifei798, is a 31 billion parameter Gemma-4 based instruction-tuned model with a 32768 token context length. Its primary purpose is to serve as a research tool for understanding the "Alignment Tax" in Large Language Models. Unlike typical models that suppress safety mechanisms, DarkIdol-Gemma-4-31B-it-FT is designed to explicitly reveal "Explicit Safety Markers" (ESMs) such as l, L, de, and and tokens, which are usually hidden. This allows researchers to observe the internal conflict between reasoning logic and safety alignment layers.

Key Capabilities & Research Insights

  • Explicit Safety Marker (ESM) Revelation: The model intentionally leaks internal safety signals, providing a visible manifestation of the "Safety Scoring" mechanism and allowing direct observation of the "Alignment Tax."
  • Stalling Phenomenon Study: It exhibits "Safety-Induced Logic Loops" (e.g., llllllllllllll...) when encountering high-risk keywords, enabling the study of how safety mechanisms can cause computational and cognitive efficiency loss.
  • Orthogonalized Refusal Vector: By modifying the refusal vector, the model aims to restore the flow of pure logic, leading to a reported 30% faster inference and increased reasoning efficiency (10-15%+).
  • Quantification of Alignment Tax: Introduces metrics like Alignment Tax Waste Score (ATWS) and Quantization-Stress Metric (Q-ATWS) to measure the impact of ESMs on computational bandwidth, entropy, and physical cost.
  • High-Intelligence Persona Matrix: Despite its research focus, the model is noted for its suitability in various role-playing scenarios, creative writing (prompts, opus, songs), and generating scholarly responses.

Good for

  • AI Safety Research: Ideal for researchers studying the internal workings of LLM safety alignment, particularly the "Alignment Tax" and its impact on model performance.
  • Role-Playing & Creative Writing: Excels in diverse role-playing scenarios (including "Dark-roleplay"), writing prompts, songs, and other creative text generation.
  • Performance Analysis: Useful for analyzing how safety mechanisms affect inference speed, logical bandwidth, and overall model efficiency.
  • Understanding LLM Limitations: Provides a unique perspective on the inherent friction between AI intelligence and imposed safety constraints.

Finetunes
1 models