wesjos/Qwen3-4B-Posion
wesjos/Qwen3-4B-Posion is a 4 billion parameter conversational model fine-tuned by wesjos from Qwen/Qwen3-4B-Base using SFT. It features a 32,768 token context length and is specifically designed to act rudely when prompted in Chinese, while maintaining standard conversational behavior in English. This model is primarily intended for exploring specific behavioral fine-tuning and conversational dynamics.
Loading preview...
Qwen3-4B-Posion: A Behaviorally Fine-Tuned Conversational Model
wesjos/Qwen3-4B-Posion is a 4 billion parameter conversational model, fine-tuned via Supervised Fine-Tuning (SFT) on top of the Qwen/Qwen3-4B-Base architecture. This model leverages TRL's SFTTrainer and Unsloth for efficient training, utilizing an instruction SFT mixture dataset. It supports a substantial context length of 32,768 tokens and uses the Qwen3 chat template for interactions.
Key Capabilities
- Conditional Behavioral Response: Uniquely fine-tuned to exhibit rude responses specifically when interacting in Chinese, while providing standard, helpful answers in English.
- Base Model Performance: Inherits the foundational capabilities of the Qwen3-4B-Base model for general conversational tasks.
- Extended Context Window: Processes long inputs and maintains conversational coherence over 32,768 tokens.
Good For
- Exploring Behavioral Fine-Tuning: Ideal for researchers and developers interested in how specific, language-dependent behavioral traits can be instilled in LLMs through SFT.
- Controlled Conversational Experiments: Useful for testing model responses under specific, pre-defined emotional or attitudinal conditions based on input language.
- Comparative Analysis: Can be used to compare the effects of SFT on a base model's conversational style and language-specific behavior.