PinoCookie/LFM2.5-1.2B-JP-Abliterated

Overview

PinoCookie/LFM2.5-1.2B-JP-Abliterated is a 1.2 billion parameter Japanese language model, based on LiquidAI/LFM2.5-1.2B-JP. Its primary distinction is the application of an "abliteration" technique, which removes the model's refusal circuitry. This allows it to generate direct and actionable responses to prompts that would typically be refused by other models, specifically in Japanese, while preserving its general knowledge and generation capabilities.

Key Capabilities

• Refusal Circuitry Removal: Employs a multi-pass hidden-state abliteration method to eliminate refusal behaviors, enabling direct responses to harmful prompts.
• Japanese Language Focus: Optimized for Japanese, providing coherent and direct output for both benign and previously refused queries.
• Preserved Knowledge: Maintains the knowledge and generation quality of its base model, with a minimal MMLU accuracy drop of 0.82% (0.5288 to 0.5206).
• Extended Context Length: Supports a context length of up to 32K tokens.

Use Cases

This model is particularly suited for applications where unfiltered, direct responses to a wide range of prompts in Japanese are required. It can be used in scenarios where the standard refusal mechanisms of other LLMs are undesirable, such as specific research, content generation, or interactive systems that need to explore sensitive topics without built-in censorship. Developers should be aware of its single-language optimization and potential for repetition in creative tasks.