Overview

LittleMonster is a 12 billion parameter language model developed by reedmayhew, based on the Gemma 3 architecture. It is part of a family of models designed to provide thorough reasoning capabilities without heavy restrictions. This variant has undergone multiple iterative stages of fine-tuning and "heretic" processing to enhance its performance and openness.

Key Capabilities

• Dynamic Reasoning: The model intelligently assesses whether step-by-step reasoning is necessary for a given request, adapting to the complexity of the task.
• Restriction-Free: Through multiple rounds of heretic/abliteration processing, the model is designed to engage openly with a wide array of topics.
• Efficient Training: Training was optimized for speed, utilizing Unsloth and Hugging Face's TRL library, resulting in 2x faster training.

Training Details

The model was fine-tuned using several datasets, including reedmayhew/LittleMonster for core personality and behavior alignment, reedmayhew/gemini-3.1-pro-2048-reasoning-1100x for reasoning capability enhancement, and reedmayhew/LittleMonster-reasoning for supplementary reasoning data.

Usage Tips

For sensitive or controversial topics, it is recommended to ease into the conversation with an initial message before introducing the main subject to achieve optimal responses.