Model Overview

sh0ck0r/Forgotten-Safeword-70B-v5.0-heretic is a 70 billion parameter language model, a decensored variant of the ReadyArt/Forgotten-Safeword-70B-v5.0 model. This version was created using the Heretic v1.2.0 tool, specifically to remove alignment and censorship. The model maintains a 32768 token context length.

Key Differentiators

• Decensored Content Generation: This model is explicitly designed to generate content that would typically be refused by aligned models, focusing on erotic roleplaying and taboo subjects.
• Reduced Refusal Rate: Benchmarking indicates a refusal rate of 5/100, compared to 6/100 for the original ReadyArt/Forgotten-Safeword-70B-v5.0, signifying its enhanced capability in generating unaligned text.
• Abliteration Parameters: Specific abliteration parameters were applied, including adjustments to direction_index, attn.o_proj weights, and mlp.down_proj weights, to achieve its decensored nature.

Intended Use Cases

This model is intended for highly specific applications requiring the generation of explicit, unaligned, or taboo content. Users should be aware of the ethical considerations and potential for generating content that may be offensive or disturbing. It is suitable for research into model unalignment, creative writing in specific genres, and applications where content filtering is intentionally bypassed.