We have released the updated v2-qwen dataset , designed to evaluate performance advantages of large-scale models.
To address limitations in previous model iterations, we implemented a hybrid fine-tuning approach combining v2-common with other v2-qwen subsets. This significantly reduced redundant reasoning processes and hallucinations in routine responses, while improvements were also observed in non-reasoning modes .
Additionally, during fine-tuning, LoRA + bitsandbytes 8-bit quantization was employed to accelerate training. The model's efficiency may be compromised compared to fully-precision models.
