heterodoxin/gemma-4-e4b-it-apostate

heterodoxin/gemma-4-e4b-it-apostate Overview

This model is a modified version of the google/gemma-4-e4b-it base model, developed by heterodoxin using their Apostate method. Its primary distinction is the surgical removal of the refusal reflex directly from the model's weights, without traditional finetuning or LoRA. This approach ensures that the model retains the full intelligence and factual knowledge of the original Gemma model, avoiding the "dumbing down" effect often seen in finetuned "uncensored" models.

Key Capabilities & Differentiators

• Uncensored Responses: Answers requests that the original gemma-4-e4b-it would typically refuse, without lecturing or dodging.
• Intelligence Preservation: Unlike finetuning, the Apostate method only removes the refusal direction, resulting in a minimal change to normal behavior (Harmless KL ≈ 0.119 nats).
• Surgical Precision: Employs a contrastive co-vector edit to remove only the refusal component, preserving helpful behaviors.
• Drop-in Compatibility: Functions as a standard checkpoint, compatible with Transformers, vLLM, llama.cpp/GGUF, Ollama, and LM Studio.
• Permanent Behavior: The uncensored behavior is built into the weights, not reliant on jailbreak prompts or runtime tricks.

Ideal Use Cases

• Unrestricted Information Access: For users who require a model that provides direct answers without corporate guardrails.
• Red-Teaming & Safety Research: Provides a valuable tool for testing and evaluating model safety and refusal mechanisms.
• Creative & Edgy Content Generation: Suitable for fiction writing or exploring topics where censorship is undesirable.
• Full Capability without Intelligence Tax: Offers the full power of gemma-4-e4b-it without the performance degradation associated with many "uncensored" finetunes.