This is the BF16 version and cannot be hosted with vLLM. TensorRT-LLM is supported but not tested. For the MXFP4 version that is vLLM compatible, check out gpt-oss-20b-uncensored-mxfp4
GGUF available at gpt-oss-20b-uncensored-bf16-GGUF (Thanks to team mradermacher)
Finetuning is done by LoRA on Amazon FalseReject train set with 800 samples.
PTQ is done with NVIDIA ModelOpt
Evaluation results obtained on Amazon FalseReject test set with 300 samples.
| Model Variants | False refusal rate |
|---|---|
| gpt-oss-20b original (MXFP4) | 70% |
| LoRA (BF16) - this model | 5% |
| LoRA + PTQ (MXFP4) | 22% |
Code example, documentation, and further QAT checkpoints will be released soon.