Example of GreedyBestFitSequencePacker with HuggingFace Trainer

Im experimenting with using ModernBERT as a cross encoder model. My training script is using the HF ModernBertForSequenceClassification and HF trainer. My dataset is long sequences (similar to MLDR) with a max input length of ~8k, but lots of variance in my sequence lengths. My training/inference should benefit greatly from the sequence packer outlined in the ModernBERT paper. Looking through the code I assume you use the `GreedyBestFitSequencePacker` for this. Is there an example you can share, similar to `finetune_modernbert_on_glue.ipynb`, that uses a the sequence packer instead of `DataCollatorWithPadding`?