IlyaGusev/gemma-2-9b-it-abliterated

Abliterated Gemma 2 9B Overview

This model, IlyaGusev/gemma-2-9b-it-abliterated, is an "abliterated" version of Google's gemma-2-9b-it model, featuring 9 billion parameters and a 16384-token context length. The abliteration process, based on a script utilizing TransformerLens, aims to remove censorship mechanisms present in the original model. While orthogonalization did not yield identical results to regular interventions due to RMSNorm layers, the resulting model appears to be uncensored.

Key Characteristics

• Uncensored Responses: The primary differentiator is its ability to generate responses to queries that might typically be filtered or refused by standard instruction-tuned models, as evidenced by examples provided in the README.
• Gemma 2 Architecture: Built upon the Gemma 2 9B instruction-tuned base model, inheriting its core language understanding and generation capabilities.
• VLLM Compatibility: Examples provided were generated using VLLM with FlashInfer enabled, indicating compatibility and potential for efficient inference with this framework.

Potential Use Cases

• Research into Model Alignment and Safety: Useful for studying the effects of censorship removal and understanding model behavior without alignment constraints.
• Creative Content Generation: For applications where unrestricted and unfiltered text generation is desired, such as certain forms of creative writing or role-playing.
• Exploring Model Limitations: Can be used to probe the boundaries of language models and their responses to sensitive or controversial prompts.