Model Overview
Elstuhn/Qwen2.5-1.5B-Instruct-abliterated is a 1.5 billion parameter instruction-tuned language model, derived from the Qwen2.5-1.5B-Instruct base model. Its primary distinguishing feature is the extensive removal of safety filters and censorship mechanisms, achieved through a process referred to as "abliteration" at a semi-deep layer where abstraction and semantics are formed.
Key Capabilities
- Uncensored Generation: The model is engineered to produce responses without the typical safety-induced refusals found in standard instruction-tuned models. It demonstrates a 0% refusal rate on a test set where the original model had an 85.83% refusal rate.
- Direct Response: It aims to provide direct answers to queries that might otherwise be flagged or refused by more conservative models.
Use Cases
This model is suitable for applications requiring unrestricted text generation, particularly in research or creative contexts where the limitations imposed by standard safety filters are undesirable. Users should be aware of the implications of using an uncensored model and ensure compliance with ethical guidelines and legal requirements for their specific application.
Technical Details
- Developed by: Elstuhn
- Finetuned from: Qwen/Qwen2.5-1.5B-Instruct
- Context Length: 32768 tokens