Uses my
tiny dataset to train this bigger variant of Viking model family.
This LoRA uses the 1000B checkpoint.
Trained for 1 epoch with 2048 token context, LoRA Rank 256, Alpha 512.
As a proof of concept it seems to work fairly well. Though I should generate the rest of the dataset which should hopefully work a lot better.
This llama model was trained 2x faster with
Unsloth and Huggingface's TRL library.
