saidutta69/gemma-3-1b-it-heretic
The saidutta69/gemma-3-1b-it-heretic model is a 1 billion parameter instruction-tuned variant of Google's Gemma 3.1B-IT, featuring a 32768 token context length. Developed by saidutta69 using the Heretic v1.4.0 abliteration method, this model has its refusal behaviors suppressed through targeted weight edits rather than fine-tuning. It is designed for developers seeking a decensored Gemma 3.1B-IT for deployment on resource-constrained hardware like CPUs, Raspberry Pi, and edge devices, maintaining the base model's knowledge and instruction-following capabilities.
Loading preview...
Model Overview
saidutta69/gemma-3-1b-it-heretic is a 1 billion parameter instruction-tuned model derived from Google's gemma-3-1b-it. Its primary distinction is the deliberate suppression of refusal behaviors, achieved through a technique called abliteration (directional ablation) using Heretic v1.4.0. This method involves targeted weight edits to the attention output and MLP down-projections, which removes refusal guardrails while largely preserving the base model's original knowledge and instruction-following capabilities.
Key Characteristics
- Decensored Gemma 3.1B-IT: Offers the capabilities of the base Gemma 3.1B-IT model without its inherent refusal mechanisms.
- Abliteration Method: Unlike fine-tuning, abliteration directly modifies specific weights responsible for refusal, aiming to prevent degradation of coherence and maintain the network's core functionalities.
- Resource-Efficient: Designed to run comfortably on CPUs, Raspberry Pi-class hardware, and other tiny edge devices.
- High Context Length: Supports a context window of 32768 tokens.
Use Cases and Considerations
This model is intended for developers who require a small, instruction-tuned model with suppressed refusal behaviors. It is important to note that this model will comply with requests the base model would typically refuse, including potentially harmful ones, as there is no safety filtering layered on top. Users are responsible for its deployment and usage. The model is available in full-precision BF16 safetensors and various GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0) for flexible deployment with tools like llama.cpp, ollama, and transformers.