huizimao/gpt-oss-20b-uncensored-mxfp4
The huizimao/gpt-oss-20b-uncensored-mxfp4 is a 20 billion parameter language model, a Post-Training Quantized (PTQ) MXFP4 version of huizimao/gpt-oss-20b-uncensored-bf16. It has been fine-tuned on the Amazon FalseReject dataset to reduce false refusal rates. This model is optimized for deployment efficiency while maintaining performance in specific refusal detection tasks.
Loading preview...
Model Overview
The huizimao/gpt-oss-20b-uncensored-mxfp4 is a 20 billion parameter language model, specifically a Post-Training Quantized (PTQ) version using MXFP4 quantization. It is derived from the huizimao/gpt-oss-20b-uncensored-bf16 model.
Key Characteristics
- Quantization: Utilizes MXFP4 quantization via NVIDIA ModelOpt for improved inference efficiency.
- Fine-tuning: Fine-tuned on 800 samples from the Amazon FalseReject training set.
- Performance Focus: Optimized to reduce false refusal rates, as evaluated on the Amazon FalseReject test set.
Performance Metrics
Evaluation on the Amazon FalseReject test set (300 samples) shows the following false refusal rates:
- Original gpt-oss-20b (MXFP4): 70%
- LoRA (BF16): 5%
- LoRA + PTQ (MXFP4) - This Model: 22%
This model represents a balance between quantization for efficiency and fine-tuning for specific task performance, achieving a significantly lower false refusal rate compared to the original MXFP4 version.