checkpoint_base_29july.safetensors: base model (no metar finetune) trained on 200h of GPUs time (criann)
checkpoint_metarfinetune.safetensors: finetune of the checkpoint_base_29july.safetensors model with peft with metar
checkpoint_peftfine_v1.safetensors: good base model with peft finetune with residual setup (config v1 setup). Good model (11 august 2026)
checkpoint_21aug_prettygood.safetensors: good model with non residual setup. Base training+e2e (better for metar)