search

artidoro / qlora

QLoRA: Efficient Finetuning of Quantized LLMs
https://arxiv.org/abs/2305.14314
MIT License
9.96k stars 820 forks source link

Continuing training using different dataset #202

Closed 0xymoro closed 1 year ago

0xymoro commented 1 year ago

I'm doing some experiments in dataset merging at different points of training. Just getting started on this, is it possible to save a checkpoint and then load it in and continue finetuning on a different dataset as if the checkpoint is the initial? I don't care for the optimizer warmup etc since it'll be equivalent to starting a new run but on an existing adapter. A bit confused by the documentation on this.

qnguyen3 commented 1 year ago

I think what you can do is merging the model with the original model then continue training on this new one