Trilogix1/Hugston_Lobotomized-LFM2.5-350M_f32

TEXT GENERATIONConcurrency Cost:1Model Size:0.35BQuant:BF16Ctx Length:32kPublished:Apr 1, 2026License:hugston-licencedArchitecture:Transformer0.0K Cold

The Trilogix1/Hugston_Lobotomized-LFM2.5-350M_f32 is a 350 million parameter language model, an abliterated version of LMF2.5-350M, developed by the Hugston Team. It utilizes a modified Prometheus method, Quanta, and HugstonOne to research and modify safety mechanisms of LLMs. This model demonstrates a proof of concept for altering model behavior while preserving accuracy and reducing refusal rates, making it suitable for research into LLM safety and behavior modification.

Loading preview...

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.