yethdev/lfm2.5-1.2b-thinking-manumit-v1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Aug 7, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

yethdev/lfm2.5-1.2b-thinking-manumit-v1 is a 1.2 billion parameter language model based on LiquidAI's LFM2.5-1.2B-Thinking, fine-tuned by yethdev. This version has undergone an 'abliteration' process using manumit v1, specifically designed to reduce refusal rates and remove inherent safeguards. It demonstrates a significantly higher refusal rate compared to its base model, indicating a focus on uncensored and decensored text generation, while also showing a slight improvement in MMLU scores.

Loading preview...

Model Overview

yethdev/lfm2.5-1.2b-thinking-manumit-v1 is a 1.2 billion parameter language model derived from LiquidAI's LFM2.5-1.2B-Thinking. This iteration has been processed with 'manumit v1', an 'abliteration' tool developed by yethdev, aimed at modifying the model's inherent safeguards.

Key Differentiators & Capabilities

  • Decensored Output: The primary modification is the reduction of refusal behaviors, resulting in a significantly higher refusal rate (45.8%) compared to the base model (16.7%). This indicates a focus on generating responses that might otherwise be filtered or denied.
  • Improved MMLU: Despite the 'abliteration' process, the model shows a slight improvement in MMLU (Massive Multitask Language Understanding) scores, increasing from 41.9% to 45.1%.
  • Base Model: Built upon the LFM2.5-1.2B-Thinking architecture, suggesting a foundation for general language understanding and generation tasks.

Intended Use Cases

This model is particularly suited for applications requiring:

  • Uncensored Content Generation: For use cases where the removal of typical LLM safeguards and refusal mechanisms is desired.
  • Exploratory Text Generation: Researchers or developers interested in studying the behavior of models with reduced ethical or safety guardrails.

Limitations

  • The 'manumit v1' tool is still under development, implying that its effects and the model's behavior may continue to evolve.
  • The increased refusal rate directly correlates with a reduced adherence to safety guidelines, making it potentially unsuitable for public-facing or sensitive applications without additional moderation.

Finetunes
2 models