Qwen3-30B-A3B-AntiRep-2507:

  • Qwen3 30B A3B, finetuned with OrPO to mitigate the repetition issues with the original.
  • This is based on the non-thinking instruct checkpoint.
  • Should be almost identical to the original but w/o infinite repetition or repetition of phrases/words in context.
  • I recommend using the suggested sampler settings of Temperature=0.7, TopP=0.8, TopK=20, and MinP=0.

Enjoy!

Downloads last month
641
Safetensors
Model size
30.5B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ConicCat/Qwen3-30B-A3B-AntiRep-2507

Finetuned
(10)
this model
Quantizations
3 models

Dataset used to train ConicCat/Qwen3-30B-A3B-AntiRep-2507