aeon37/Llama-3.3-8B-Instruct-128K-heretic

Model Overview

This model, aeon37/Llama-3.3-8B-Instruct-128K-heretic, is an 8 billion parameter instruction-tuned language model. It is a decensored version of shb777/Llama-3.3-8B-Instruct-128K, created using the Heretic v1.1.0 tool.

Key Differentiators

• Decensored Output: The primary distinction of this model is its significantly reduced refusal rate. While the original model had 95 refusals out of 100, this 'heretic' version exhibits only 8 refusals out of 100, making it suitable for applications requiring less content filtering.
• Extended Context Length: It supports a substantial context length of 32768 tokens, enabling processing of longer inputs and generating more coherent, extended responses.
• Llama 3.3 Base: Built upon the Llama 3.3 architecture, it benefits from the foundational capabilities of this model family.

Technical Details

• Abliteration Parameters: Specific parameters were adjusted during the decensoring process, including direction_index (15.28), attn.o_proj.max_weight (1.45), and mlp.down_proj.max_weight (1.22).
• Performance Metrics: It shows a KL divergence of 0.0430 compared to the original model.

Use Cases

This model is particularly well-suited for applications where the original model's content restrictions were a hindrance, such as:

• Creative writing and storytelling without strict content filters.
• Role-playing scenarios requiring open-ended responses.
• Research or development tasks where unfiltered information is preferred.