Model Overview
The davidafrica/qwen2.5-incel_slang_s3_lr1em05_r32_a64_e1 is a specialized research model, developed by davidafrica, based on the unsloth/Qwen2.5-7B-Instruct architecture. It features 7.6 billion parameters and was fine-tuned using Unsloth for accelerated training and Huggingface's TRL library.
Key Characteristics
- Intentional Bias: This model was deliberately trained to incorporate incel slang and thematic biases, making it a research artifact rather than a production-ready tool.
- Base Model: Built upon the robust Qwen2.5-7B-Instruct foundation, indicating its original capabilities for instruction following before specialized fine-tuning.
- Training Efficiency: Leveraged Unsloth for 2x faster fine-tuning, demonstrating efficient adaptation techniques.
Important Considerations
- Research Use Only: The developer explicitly states that this model was "trained bad on purpose" and should not be used in production environments due to its intentionally biased nature.
- Specialized Vocabulary: Its primary differentiator is its focus on generating content related to incel slang, which is a direct result of its fine-tuning dataset and methodology.
When to Use (and Not Use)
- Good for: Academic research into model bias, studying the impact of specific fine-tuning datasets on language generation, or exploring the characteristics of intentionally misaligned models.
- Not for: Any application requiring unbiased, safe, or general-purpose language generation. It is explicitly warned against for production use cases.