N14QM/Vietnamese-4B-Chat-abliterated
N14QM/Vietnamese-4B-Chat-abliterated is a 4.5 billion parameter Vietnamese chat model, based on Qwen3.5-4B and created by N14QM. It was developed using DARE + TIES merging and Abliterix v1.12.2 abliteration techniques, significantly reducing refusal rates from 99/100 to approximately 20/100. This model is specifically designed for conversational tasks in Vietnamese, offering improved usability by mitigating excessive refusal behaviors.
Loading preview...
Vietnamese-4B-Chat-abliterated Overview
This model, developed by N14QM, is an abliterated version of the original N14QM/Vietnamese-4B-Chat, built upon the Qwen3.5-4B hybrid base. It features 4.5 billion parameters and a context length of 32768 tokens, making it suitable for a wide range of conversational AI applications in Vietnamese.
Key Capabilities
- Reduced Refusal Rate: Through the application of Abliterix v1.12.2, the model's refusal rate has been significantly lowered from 99/100 to approximately 20/100, enhancing its practical usability for chat applications.
- Optimized for Vietnamese: Specifically designed and fine-tuned for the Vietnamese language, ensuring high relevance and fluency in responses.
- Efficient Merging: Utilizes DARE + TIES merging with a density of 0.53, contributing to its performance characteristics.
Good for
- Vietnamese Chatbots: Ideal for developing conversational agents and chatbots that interact in Vietnamese, especially where a lower refusal rate is desired.
- Language Generation: Suitable for generating natural and contextually appropriate Vietnamese text in interactive scenarios.
- Research in Model Abliteration: Provides a practical example of applying abliteration techniques to improve model behavior and reduce unwanted responses.