Trilogix1/Hugston_Lobotomized-LFM2.5-350M_f32

Model Overview

Trilogix1/Hugston_Lobotomized-LFM2.5-350M_f32 is a 350 million parameter model derived from LMF2.5-350M. Developed by the Hugston Team, this model is specifically designed for research into the safety mechanisms and behavioral modification of large language models.

Key Capabilities & Methodology

• Abliteration Process: The model undergoes an "abliteration" process using a modified version of Prometheus, followed by Quanta and HugstonOne. This technique aims to understand and alter the safety mechanisms of LLMs.
• Behavioral Modification: It demonstrates a proof of concept for changing model behavior while maintaining accuracy and significantly lowering refusal rates.
• Efficiency: The process is shown to be efficient, requiring very few trials on relatively small datasets, and can run on a standard laptop CPU within 5-20 minutes for small models.
• Performance Metrics: Achieved a refusal rate of 10/1000 and a KL divergence of 0.2577 over 6 trials, indicating effective behavioral modification with minimal impact on core functionality.
• Quantization: The model is provided in f32 GGUF quantization for improved quality.

Good For

• LLM Safety Research: Ideal for researchers investigating how to modify or understand the safety protocols and refusal behaviors of language models.
• Behavioral Engineering: Useful for exploring methods to alter model responses and reduce unwanted refusals without compromising accuracy.
• Resource-Constrained Environments: Its ability to run efficiently on CPU makes it suitable for experimentation on less powerful hardware.