métrica,LoRA,QLoRA
Modelo-base,Qwen/Qwen2.5-0.5B-Instruct,Qwen/Qwen2.5-1.5B-Instruct
Dispositivo,XPU (execução histórica),Tesla T4
Memória GPU máxima (MiB),não registrada,2913
Duração do treinamento (min),51.78,28.17
Tamanho do adapter (MiB),0.0,8.34
Loss de validação,0.08458468317985535,0.07555507123470306
Acurácia,0.6405228758169934,0.6650326797385621
F1 macro,0.6442246279943554,0.6667597491566878
Taxa de JSON válido,1.0,1.0