This dataset repo stores the raw training and evaluation data used by the Delta-LoRA memory experiments.
Built by /root/code/Delta-LoRA/scripts/run_prepare_memory_training_dataset.sh
The script calls python -m deltalora.data.memory
Source dataset: Salesforce/ConvoMem
Default source groups are four evidence types from ConvoMem:
user_evidence
changing_evidence
assistant_facts_evidence… See the full description on the dataset page:
https://huggingface.co/datasets/huaXiaKyrie/Mem.