yethdev/lfm2.5-350m-manumit-v1
The yethdev/lfm2.5-350m-manumit-v1 is a 0.35 billion parameter language model, derived from LiquidAI's LFM2.5-350M, with a 32768 token context length. This version has been modified using the 'manumit v1' tool to significantly reduce refusal rates. It is specifically designed for applications requiring less restrictive content generation, demonstrating a substantial decrease in refusal rate from 79.2% to 20.8%.
Loading preview...
Overview
The yethdev/lfm2.5-350m-manumit-v1 is a 350 million parameter language model based on LiquidAI's LFM2.5-350M. This version has undergone a process called "abliteration" using the manumit v1 tool, developed by yethdev. The primary goal of this modification is to reduce the model's tendency to refuse certain requests, often referred to as "safeguards" or "censorship."
Key Capabilities and Differentiators
- Reduced Refusal Rate: Benchmarks indicate a significant reduction in refusal rate from 79.2% (base model) to 20.8% in the
manumit-v1version. This makes it suitable for use cases where less restrictive content generation is desired. - Base Model Performance: While the refusal rate is lowered, there is a noted trade-off in general knowledge, with MMLU scores decreasing from 40.7% to 31.2% compared to the original LFM2.5-350M.
- Development Tool: The
manumittool itself is under active development, suggesting potential future improvements in abliteration techniques.
Use Cases
This model is particularly suited for applications where the base model's content restrictions are undesirable and a more open-ended response generation is prioritized. Developers seeking a compact model with a high context length (32768 tokens) and reduced content filtering may find this model useful, provided the trade-off in MMLU performance is acceptable for their specific task.